Zheng Zhu

Yi Wei, Linqing Zhao, Wenzhao Zheng et al. · tsinghua

3D scene understanding plays a vital role in vision-based autonomous driving. While most existing methods focus on 3D object detection, they have difficulty describing real-world objects of arbitrary shapes and infinite classes. Towards a more comprehensive perception of a 3D scene, in this paper, we propose a SurroundOcc method to predict the 3D occupancy with multi-camera images. We first extract multi-scale features for each image and adopt spatial 2D-3D attention to lift them to the 3D volume space. Then we apply 3D convolutions to progressively upsample the volume features and impose supervision on multiple levels. To obtain dense occupancy prediction, we design a pipeline to generate dense occupancy ground truth without expansive occupancy annotations. Specifically, we fuse multi-frame LiDAR scans of dynamic objects and static scenes separately. Then we adopt Poisson Reconstruction to fill the holes and voxelize the mesh to get dense occupancy labels. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our method. Code and dataset are available at https://github.com/weiyithu/SurroundOcc

Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng et al. · tsinghua

In this paper, we present BEVerse, a unified framework for 3D perception and prediction based on multi-camera systems. Unlike existing studies focusing on the improvement of single-task approaches, BEVerse features in producing spatio-temporal Birds-Eye-View (BEV) representations from multi-camera videos and jointly reasoning about multiple tasks for vision-centric autonomous driving. Specifically, BEVerse first performs shared feature extraction and lifting to generate 4D BEV representations from multi-timestamp and multi-view images. After the ego-motion alignment, the spatio-temporal encoder is utilized for further feature extraction in BEV. Finally, multiple task decoders are attached for joint reasoning and prediction. Within the decoders, we propose the grid sampler to generate BEV features with different ranges and granularities for different tasks. Also, we design the method of iterative flow for memory-efficient future prediction. We show that the temporal information improves 3D object detection and semantic map construction, while the multi-task learning can implicitly benefit motion prediction. With extensive experiments on the nuScenes dataset, we show that the multi-task BEVerse outperforms existing single-task methods on 3D object detection, semantic map construction, and motion prediction. Compared with the sequential paradigm, BEVerse also favors in significantly improved efficiency. The code and trained models will be released at https://github.com/zhangyp15/BEVerse.

Shuai Shen, Wanhua Li, Zheng Zhu et al. · tsinghua

Talking head synthesis is an emerging technology with wide applications in film dubbing, virtual avatars and online education. Recent NeRF-based methods generate more natural talking videos, as they better capture the 3D structural information of faces. However, a specific model needs to be trained for each identity with a large dataset. In this paper, we propose Dynamic Facial Radiance Fields (DFRF) for few-shot talking head synthesis, which can rapidly generalize to an unseen identity with few training data. Different from the existing NeRF-based methods which directly encode the 3D geometry and appearance of a specific person into the network, our DFRF conditions face radiance field on 2D appearance images to learn the face prior. Thus the facial radiance field can be flexibly adjusted to the new identity with few reference images. Additionally, for better modeling of the facial deformations, we propose a differentiable face warping module conditioned on audio signals to deform all reference images to the query space. Extensive experiments show that with only tens of seconds of training clip available, our proposed DFRF can synthesize natural and high-quality audio-driven talking head videos for novel identities with only 40k iterations. We highly recommend readers view our supplementary video for intuitive comparisons. Code is available in https://sstzal.github.io/DFRF/.

Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu et al. · tsinghua

3D object detection with surrounding cameras has been a promising direction for autonomous driving. In this paper, we present SimMOD, a Simple baseline for Multi-camera Object Detection, to solve the problem. To incorporate multi-view information as well as build upon previous efforts on monocular 3D object detection, the framework is built on sample-wise object proposals and designed to work in a two-stage manner. First, we extract multi-scale features and generate the perspective object proposals on each monocular image. Second, the multi-view proposals are aggregated and then iteratively refined with multi-view and multi-scale visual features in the DETR3D-style. The refined proposals are end-to-end decoded into the detection results. To further boost the performance, we incorporate the auxiliary branches alongside the proposal generation to enhance the feature learning. Also, we design the methods of target filtering and teacher forcing to promote the consistency of two-stage training. We conduct extensive experiments on the 3D object detection benchmark of nuScenes to demonstrate the effectiveness of SimMOD and achieve new state-of-the-art performance. Code will be available at https://github.com/zhangyp15/SimMOD.

Chengkun Wang, Wenzhao Zheng, Zheng Zhu et al. · tsinghua

This paper proposes an introspective deep metric learning (IDML) framework for uncertainty-aware comparisons of images. Conventional deep metric learning methods focus on learning a discriminative embedding to describe the semantic features of images, which ignore the existence of uncertainty in each image resulting from noise or semantic ambiguity. Training without awareness of these uncertainties causes the model to overfit the annotated labels during training and produce unsatisfactory judgments during inference. Motivated by this, we argue that a good similarity model should consider the semantic discrepancies with awareness of the uncertainty to better deal with ambiguous images for more robust training. To achieve this, we propose to represent an image using not only a semantic embedding but also an accompanying uncertainty embedding, which describes the semantic characteristics and ambiguity of an image, respectively. We further propose an introspective similarity metric to make similarity judgments between images considering both their semantic differences and ambiguities. The gradient analysis of the proposed metric shows that it enables the model to learn at an adaptive and slower pace to deal with the uncertainty during training. The proposed IDML framework improves the performance of deep metric learning through uncertainty modeling and attains state-of-the-art results on the widely used CUB-200-2011, Cars196, and Stanford Online Products datasets for image retrieval and clustering. We further provide an in-depth analysis of our framework to demonstrate the effectiveness and reliability of IDML. Code: https://github.com/wzzheng/IDML.

Han Xiao, Wenzhao Zheng, Zheng Zhu et al. · tsinghua

Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphoria16/TL-Align.

Zhang Youmin, Guo Xianda, Poggi Matteo et al.

Given sparse depths and the corresponding RGB images, depth completion aims at spatially propagating the sparse measurements throughout the whole image to get a dense depth prediction. Despite the tremendous progress of deep-learning-based depth completion methods, the locality of the convolutional layer or graph model makes it hard for the network to model the long-range relationship between pixels. While recent fully Transformer-based architecture has reported encouraging results with the global receptive field, the performance and efficiency gaps to the well-developed CNN models still exist because of its deteriorative local feature details. This paper proposes a Joint Convolutional Attention and Transformer block (JCAT), which deeply couples the convolutional attention layer and Vision Transformer into one block, as the basic unit to construct our depth completion model in a pyramidal structure. This hybrid architecture naturally benefits both the local connectivity of convolutions and the global context of the Transformer in one single model. As a result, our CompletionFormer outperforms state-of-the-art CNNs-based methods on the outdoor KITTI Depth Completion benchmark and indoor NYUv2 dataset, achieving significantly higher efficiency (nearly 1/3 FLOPs) compared to pure Transformer-based methods. Code is available at \url{https://github.com/youmi-zym/CompletionFormer}.

Yi Wei, Linqing Zhao, Wenzhao Zheng et al. · tsinghua

Depth estimation from images serves as the fundamental step of 3D perception for autonomous driving and is an economical alternative to expensive depth sensors like LiDAR. The temporal photometric constraints enables self-supervised depth estimation without labels, further facilitating its application. However, most existing methods predict the depth solely based on each monocular image and ignore the correlations among multiple surrounding cameras, which are typically available for modern self-driving vehicles. In this paper, we propose a SurroundDepth method to incorporate the information from multiple surrounding views to predict depth maps across cameras. Specifically, we employ a joint network to process all the surrounding views and propose a cross-view transformer to effectively fuse the information from multiple views. We apply cross-view self-attention to efficiently enable the global interactions between multi-camera feature maps. Different from self-supervised monocular depth estimation, we are able to predict real-world scales given multi-camera extrinsic matrices. To achieve this goal, we adopt the two-frame structure-from-motion to extract scale-aware pseudo depths to pretrain the models. Further, instead of predicting the ego-motion of each individual camera, we estimate a universal ego-motion of the vehicle and transfer it to each view to achieve multi-view ego-motion consistency. In experiments, our method achieves the state-of-the-art performance on the challenging multi-camera depth estimation datasets DDAD and nuScenes.

Zheng Zhu, Guan Huang, Jiankang Deng et al. · tsinghua

Face benchmarks empower the research community to train and evaluate high-performance face recognition systems. In this paper, we contribute a new million-scale recognition benchmark, containing uncurated 4M identities/260M faces (WebFace260M) and cleaned 2M identities/42M faces (WebFace42M) training data, as well as an elaborately designed time-constrained evaluation protocol. Firstly, we collect 4M name lists and download 260M faces from the Internet. Then, a Cleaning Automatically utilizing Self-Training (CAST) pipeline is devised to purify the tremendous WebFace260M, which is efficient and scalable. To the best of our knowledge, the cleaned WebFace42M is the largest public face recognition training set and we expect to close the data gap between academia and industry. Referring to practical deployments, Face Recognition Under Inference Time conStraint (FRUITS) protocol and a new test set with rich attributes are constructed. Besides, we gather a large-scale masked face sub-set for biometrics assessment under COVID-19. For a comprehensive evaluation of face matchers, three recognition tasks are performed under standard, masked and unbiased settings, respectively. Equipped with this benchmark, we delve into million-scale face recognition problems. A distributed framework is developed to train face recognition models efficiently without tampering with the performance. Enabled by WebFace42M, we reduce 40% failure rate on the challenging IJB-C set and rank 3rd among 430 entries on NIST-FRVT. Even 10% data (WebFace4M) shows superior performance compared with the public training sets. Furthermore, comprehensive baselines are established under the FRUITS-100/500/1000 milliseconds protocols. The proposed benchmark shows enormous potential on standard, masked and unbiased face recognition scenarios. Our WebFace260M website is https://www.face-benchmark.org.

Yunpeng Zhang, Zheng Zhu, Dalong Du

The vision-based perception for autonomous driving has undergone a transformation from the bird-eye-view (BEV) representations to the 3D semantic occupancy. Compared with the BEV planes, the 3D semantic occupancy further provides structural information along the vertical direction. This paper presents OccFormer, a dual-path transformer network to effectively process the 3D volume for semantic occupancy prediction. OccFormer achieves a long-range, dynamic, and efficient encoding of the camera-generated 3D voxel features. It is obtained by decomposing the heavy 3D processing into the local and global transformer pathways along the horizontal plane. For the occupancy decoder, we adapt the vanilla Mask2Former for 3D semantic occupancy by proposing preserve-pooling and class-guided sampling, which notably mitigate the sparsity and class imbalance. Experimental results demonstrate that OccFormer significantly outperforms existing methods for semantic scene completion on SemanticKITTI dataset and for LiDAR semantic segmentation on nuScenes dataset. Code is available at \url{https://github.com/zhangyp15/OccFormer}.

Xiaofeng Wang, Zheng Zhu, Wenbo Xu et al.

Semantic occupancy perception is essential for autonomous driving, as automated vehicles require a fine-grained perception of the 3D urban structures. However, existing relevant benchmarks lack diversity in urban scenes, and they only evaluate front-view predictions. Towards a comprehensive benchmarking of surrounding perception algorithms, we propose OpenOccupancy, which is the first surrounding semantic occupancy perception benchmark. In the OpenOccupancy benchmark, we extend the large-scale nuScenes dataset with dense semantic occupancy annotations. Previous annotations rely on LiDAR points superimposition, where some occupancy labels are missed due to sparse LiDAR channels. To mitigate the problem, we introduce the Augmenting And Purifying (AAP) pipeline to ~2x densify the annotations, where ~4000 human hours are involved in the labeling process. Besides, camera-based, LiDAR-based and multi-modal baselines are established for the OpenOccupancy benchmark. Furthermore, considering the complexity of surrounding occupancy perception lies in the computational burden of high-resolution 3D predictions, we propose the Cascade Occupancy Network (CONet) to refine the coarse prediction, which relatively enhances the performance by ~30% than the baseline. We hope the OpenOccupancy benchmark will boost the development of surrounding occupancy perception algorithms.

Jiayu Zou, Junrui Xiao, Zheng Zhu et al.

Autonomous driving requires accurate and detailed Bird's Eye View (BEV) semantic segmentation for decision making, which is one of the most challenging tasks for high-level scene perception. Feature transformation from frontal view to BEV is the pivotal technology for BEV semantic segmentation. Existing works can be roughly classified into two categories, i.e., Camera model-Based Feature Transformation (CBFT) and Camera model-Free Feature Transformation (CFFT). In this paper, we empirically analyze the vital differences between CBFT and CFFT. The former transforms features based on the flat-world assumption, which may cause distortion of regions lying above the ground plane. The latter is limited in the segmentation performance due to the absence of geometric priors and time-consuming computation. In order to reap the benefits and avoid the drawbacks of CBFT and CFFT, we propose a novel framework with a Hybrid Feature Transformation module (HFT). Specifically, we decouple the feature maps produced by HFT for estimating the layout of outdoor scenes in BEV. Furthermore, we design a mutual learning scheme to augment hybrid transformation by applying feature mimicking. Notably, extensive experiments demonstrate that with negligible extra overhead, HFT achieves a relative improvement of 13.3% on the Argoverse dataset and 16.8% on the KITTI 3D Object datasets compared to the best-performing existing method. The codes are available at https://github.com/JiayuZou2020/HFT.

Xiaofeng Wang, Zheng Zhu, Guan Huang et al.

Self-supervised monocular methods can efficiently learn depth information of weakly textured surfaces or reflective objects. However, the depth accuracy is limited due to the inherent ambiguity in monocular geometric modeling. In contrast, multi-frame depth estimation methods improve the depth accuracy thanks to the success of Multi-View Stereo (MVS), which directly makes use of geometric constraints. Unfortunately, MVS often suffers from texture-less regions, non-Lambertian surfaces, and moving objects, especially in real-world video sequences without known camera motion and depth supervision. Therefore, we propose MOVEDepth, which exploits the MOnocular cues and VElocity guidance to improve multi-frame Depth learning. Unlike existing methods that enforce consistency between MVS depth and monocular depth, MOVEDepth boosts multi-frame depth learning by directly addressing the inherent problems of MVS. The key of our approach is to utilize monocular depth as a geometric priority to construct MVS cost volume, and adjust depth candidates of cost volume under the guidance of predicted camera velocity. We further fuse monocular depth and MVS depth by learning uncertainty in the cost volume, which results in a robust depth estimation against ambiguity in multi-view geometry. Extensive experiments show MOVEDepth achieves state-of-the-art performance: Compared with Monodepth2 and PackNet, our method relatively improves the depth accuracy by 20\% and 19.8\% on the KITTI benchmark. MOVEDepth also generalizes to the more challenging DDAD benchmark, relatively outperforming ManyDepth by 7.2\%. The code is available at https://github.com/JeffWang987/MOVEDepth.

Xiaofeng Wang, Zheng Zhu, Yunpeng Zhang et al.

In recent years, vision-centric perception has flourished in various autonomous driving tasks, including 3D detection, semantic map construction, motion forecasting, and depth estimation. Nevertheless, the latency of vision-centric approaches is too high for practical deployment (e.g., most camera-based 3D detectors have a runtime greater than 300ms). To bridge the gap between ideal research and real-world applications, it is necessary to quantify the trade-off between performance and efficiency. Traditionally, autonomous-driving perception benchmarks perform the offline evaluation, neglecting the inference time delay. To mitigate the problem, we propose the Autonomous-driving StreAming Perception (ASAP) benchmark, which is the first benchmark to evaluate the online performance of vision-centric perception in autonomous driving. On the basis of the 2Hz annotated nuScenes dataset, we first propose an annotation-extending pipeline to generate high-frame-rate labels for the 12Hz raw images. Referring to the practical deployment, the Streaming Perception Under constRained-computation (SPUR) evaluation protocol is further constructed, where the 12Hz inputs are utilized for streaming evaluation under the constraints of different computational resources. In the ASAP benchmark, comprehensive experiment results reveal that the model rank alters under different constraints, suggesting that the model latency and computation budget should be considered as design choices to optimize the practical deployment. To facilitate further research, we establish baselines for camera-based streaming 3D detection, which consistently enhance the streaming performance across various hardware. ASAP project page: https://github.com/JeffWang987/ASAP.

Xianda Guo, Zheng Zhu, Tian Yang et al.

Gait benchmarks empower the research community to train and evaluate high-performance gait recognition systems. Even though growing efforts have been devoted to cross-view recognition, academia is restricted by current existing databases captured in the controlled environment. In this paper, we contribute a new benchmark and strong baseline for Gait REcognition in the Wild (GREW). The GREW dataset is constructed from natural videos, which contain hundreds of cameras and thousands of hours of streams in open systems. With tremendous manual annotations, the GREW consists of 26K identities and 128K sequences with rich attributes for unconstrained gait recognition. Moreover, we add a distractor set of over 233K sequences, making it more suitable for real-world applications. Compared with prevailing predefined cross-view datasets, the GREW has diverse and practical view variations, as well as more naturally challenging factors. To the best of our knowledge, this is the first large-scale dataset for gait recognition in the wild. Equipped with this benchmark, we dissect the unconstrained gait recognition problem, where representative appearance-based and model-based methods are explored. The proposed GREW benchmark proves to be essential for both training and evaluating gait recognizers in unconstrained scenarios. In addition, we propose the Single Path One-Shot neural architecture search with uniform sampling for Gait recognition, named SPOSGait, which is the first NAS-based gait recognition model. In experiments, SPOSGait achieves state-of-the-art performance on the CASIA-B, OU-MVLP, Gait3D, and GREW benchmarks, outperforming existing approaches by a large margin. The code will be released at https://github.com/XiandaGuo/SPOSGait.

Chaoqiang Zhao, Youmin Zhang, Matteo Poggi et al.

Self-supervised monocular depth estimation is an attractive solution that does not require hard-to-source depth labels for training. Convolutional neural networks (CNNs) have recently achieved great success in this task. However, their limited receptive field constrains existing network architectures to reason only locally, dampening the effectiveness of the self-supervised paradigm. In the light of the recent successes achieved by Vision Transformers (ViTs), we propose MonoViT, a brand-new framework combining the global reasoning enabled by ViT models with the flexibility of self-supervised monocular depth estimation. By combining plain convolutions with Transformer blocks, our model can reason locally and globally, yielding depth prediction at a higher level of detail and accuracy, allowing MonoViT to achieve state-of-the-art performance on the established KITTI dataset. Moreover, MonoViT proves its superior generalization capacities on other datasets such as Make3D and DrivingStereo.

Kai Wang, Jianyang Gu, Daquan Zhou et al.

Dataset distillation reduces the network training cost by synthesizing small and informative datasets from large-scale ones. Despite the success of the recent dataset distillation algorithms, three drawbacks still limit their wider application: i). the synthetic images perform poorly on large architectures; ii). they need to be re-optimized when the distillation ratio changes; iii). the limited diversity restricts the performance when the distillation ratio is large. In this paper, we propose a novel distillation scheme to \textbf{D}istill information of large train sets \textbf{i}nto generative \textbf{M}odels, named DiM. Specifically, DiM learns to use a generative model to store the information of the target dataset. During the distillation phase, we minimize the differences in logits predicted by a models pool between real and generated images. At the deployment stage, the generative model synthesizes various training samples from random noises on the fly. Due to the simple yet effective designs, the trained DiM can be directly applied to different distillation ratios and large architectures without extra cost. We validate the proposed DiM across 4 datasets and achieve state-of-the-art results on all of them. To the best of our knowledge, we are the first to achieve higher accuracy on complex architectures than simple ones, such as 75.1\% with ResNet-18 and 72.6\% with ConvNet-3 on ten images per class of CIFAR-10. Besides, DiM outperforms previous methods with 10\% $\sim$ 22\% when images per class are 1 and 10 on the SVHN dataset.

Xue-Feng Zhu, Tianyang Xu, Jian Zhao et al.

Unmanned Aerial Vehicles (UAVs) have been widely used in many areas, including transportation, surveillance, and military. However, their potential for safety and privacy violations is an increasing issue and highly limits their broader applications, underscoring the critical importance of UAV perception and defense (anti-UAV). Still, previous works have simplified such an anti-UAV task as a tracking problem, where the prior information of UAVs is always provided; such a scheme fails in real-world anti-UAV tasks (i.e. complex scenes, indeterminate-appear and -reappear UAVs, and real-time UAV surveillance). In this paper, we first formulate a new and practical anti-UAV problem featuring the UAVs perception in complex scenes without prior UAVs information. To benchmark such a challenging task, we propose the largest UAV dataset dubbed AntiUAV600 and a new evaluation metric. The AntiUAV600 comprises 600 video sequences of challenging scenes with random, fast, and small-scale UAVs, with over 723K thermal infrared frames densely annotated with bounding boxes. Finally, we develop a novel anti-UAV approach via an evidential collaboration of global UAVs detection and local UAVs tracking, which effectively tackles the proposed problem and can serve as a strong baseline for future research. Extensive experiments show our method outperforms SOTA approaches and validate the ability of AntiUAV600 to enhance UAV perception performance due to its large scale and complexity. Our dataset, pretrained models, and source codes will be released publically.

Xiaofeng Wang, Zheng Zhu, Guan Huang et al.

World models, especially in autonomous driving, are trending and drawing extensive attention due to their capacity for comprehending driving environments. The established world model holds immense potential for the generation of high-quality driving videos, and driving policies for safe maneuvering. However, a critical limitation in relevant research lies in its predominant focus on gaming environments or simulated settings, thereby lacking the representation of real-world driving scenarios. Therefore, we introduce DriveDreamer, a pioneering world model entirely derived from real-world driving scenarios. Regarding that modeling the world in intricate driving scenes entails an overwhelming search space, we propose harnessing the powerful diffusion model to construct a comprehensive representation of the complex environment. Furthermore, we introduce a two-stage training pipeline. In the initial phase, DriveDreamer acquires a deep understanding of structured traffic constraints, while the subsequent stage equips it with the ability to anticipate future states. The proposed DriveDreamer is the first world model established from real-world driving scenarios. We instantiate DriveDreamer on the challenging nuScenes benchmark, and extensive experiments verify that DriveDreamer empowers precise, controllable video generation that faithfully captures the structural constraints of real-world traffic scenarios. Additionally, DriveDreamer enables the generation of realistic and reasonable driving policies, opening avenues for interaction and practical applications.

Boyu Zhang, Wenbo Xu, Zheng Zhu et al.

Representing and synthesizing novel views in real-world dynamic scenes from casual monocular videos is a long-standing problem. Existing solutions typically approach dynamic scenes by applying geometry techniques or utilizing temporal information between several adjacent frames without considering the underlying background distribution in the entire scene or the transmittance over the ray dimension, limiting their performance on static and occlusion areas. Our approach $\textbf{D}$istribution-$\textbf{D}$riven neural radiance fields offers high-quality view synthesis and a 3D solution to $\textbf{D}$etach the background from the entire $\textbf{D}$ynamic scene, which is called $\text{D}^4$NeRF. Specifically, it employs a neural representation to capture the scene distribution in the static background and a 6D-input NeRF to represent dynamic objects, respectively. Each ray sample is given an additional occlusion weight to indicate the transmittance lying in the static and dynamic components. We evaluate $\text{D}^4$NeRF on public dynamic scenes and our urban driving scenes acquired from an autonomous-driving dataset. Extensive experiments demonstrate that our approach outperforms previous methods in rendering texture details and motion areas while also producing a clean static background. Our code will be released at https://github.com/Luciferbobo/D4NeRF.

Xiaofeng Wang, Zheng Zhu, Fangbo Qin et al.

Learning-based Multi-View Stereo (MVS) methods warp source images into the reference camera frustum to form 3D volumes, which are fused as a cost volume to be regularized by subsequent networks. The fusing step plays a vital role in bridging 2D semantics and 3D spatial associations. However, previous methods utilize extra networks to learn 2D information as fusing cues, underusing 3D spatial correlations and bringing additional computation costs. Therefore, we present MVSTER, which leverages the proposed epipolar Transformer to learn both 2D semantics and 3D spatial associations efficiently. Specifically, the epipolar Transformer utilizes a detachable monocular depth estimator to enhance 2D semantics and uses cross-attention to construct data-dependent 3D associations along epipolar line. Additionally, MVSTER is built in a cascade structure, where entropy-regularized optimal transport is leveraged to propagate finer depth estimations in each stage. Extensive experiments show MVSTER achieves state-of-the-art reconstruction performance with significantly higher efficiency: Compared with MVSNet and CasMVSNet, our MVSTER achieves 34% and 14% relative improvements on the DTU benchmark, with 80% and 51% relative reductions in running time. MVSTER also ranks first on Tanks&Temples-Advanced among all published works. Code is released at https://github.com/JeffWang987.

Yanqing Liu, Jianyang Gu, Kai Wang et al.

Dataset distillation plays a crucial role in creating compact datasets with similar training performance compared with original large-scale ones. This is essential for addressing the challenges of data storage and training costs. Prevalent methods facilitate knowledge transfer by matching the gradients, embedding distributions, or training trajectories of synthetic images with those of the sampled original images. Although there are various matching objectives, currently the strategy for selecting original images is limited to naive random sampling. We argue that random sampling overlooks the evenness of the selected sample distribution, which may result in noisy or biased matching targets. Besides, the sample diversity is also not constrained by random sampling. Additionally, current methods predominantly focus on single-dimensional matching, where information is not fully utilized. To address these challenges, we propose a novel matching strategy called Dataset Distillation by Bidirectional REpresentAtive Matching (DREAM+), which selects representative original images for bidirectional matching. DREAM+ is applicable to a variety of mainstream dataset distillation frameworks and significantly reduces the number of distillation iterations by more than 15 times without affecting performance. Given sufficient training time, DREAM+ can further improve the performance and achieve state-of-the-art results. We have released the code at github.com/NUS-HPC-AI-Lab/DREAM+.

Xianda Guo, Wenjie Yuan, Yunpeng Zhang et al.

Depth estimation has been widely studied and serves as the fundamental step of 3D perception for robotics and autonomous driving. Though significant progress has been made in monocular depth estimation in the past decades, these attempts are mainly conducted on the KITTI benchmark with only front-view cameras, which ignores the correlations across surround-view cameras. In this paper, we propose an Adjacent-View Transformer for Supervised Surround-view Depth estimation (AVT-SSDepth), to jointly predict the depth maps across multiple surrounding cameras. Specifically, we employ a global-to-local feature extraction module that combines CNN with transformer layers for enriched representations. Further, the adjacent-view attention mechanism is proposed to enable the intra-view and inter-view feature propagation. The former is achieved by the self-attention module within each view, while the latter is realized by the adjacent attention module, which computes the attention across multi-cameras to exchange the multi-scale representations across surroundview feature maps. In addition, AVT-SSDepth has strong crossdataset generalization. Extensive experiments show that our method achieves superior performance over existing state-ofthe-art methods on both DDAD and nuScenes datasets. Code is available at https://github.com/XiandaGuo/SSDepth.

Kai Wang, Bo Zhao, Xiangyu Peng et al.

Face recognition, as one of the most successful applications in artificial intelligence, has been widely used in security, administration, advertising, and healthcare. However, the privacy issues of public face datasets have attracted increasing attention in recent years. Previous works simply mask most areas of faces or synthesize samples using generative models to construct privacy-preserving face datasets, which overlooks the trade-off between privacy protection and data utility. In this paper, we propose a novel framework FaceMAE, where the face privacy and recognition performance are considered simultaneously. Firstly, randomly masked face images are used to train the reconstruction module in FaceMAE. We tailor the instance relation matching (IRM) module to minimize the distribution gap between real faces and FaceMAE reconstructed ones. During the deployment phase, we use trained FaceMAE to reconstruct images from masked faces of unseen identities without extra training. The risk of privacy leakage is measured based on face retrieval between reconstructed and original datasets. Experiments prove that the identities of reconstructed images are difficult to be retrieved. We also perform sufficient privacy-preserving face recognition on several public face datasets (i.e. CASIA-WebFace and WebFace260M). Compared to previous state of the arts, FaceMAE consistently \textbf{reduces at least 50\% error rate} on LFW, CFP-FP and AgeDB.

Jianqiang Xia, DianXi Shi, Ke Song et al.

Most existing RGB-T tracking networks extract modality features in a separate manner, which lacks interaction and mutual guidance between modalities. This limits the network's ability to adapt to the diverse dual-modality appearances of targets and the dynamic relationships between the modalities. Additionally, the three-stage fusion tracking paradigm followed by these networks significantly restricts the tracking speed. To overcome these problems, we propose a unified single-stage Transformer RGB-T tracking network, namely USTrack, which unifies the above three stages into a single ViT (Vision Transformer) backbone with a dual embedding layer through self-attention mechanism. With this structure, the network can extract fusion features of the template and search region under the mutual interaction of modalities. Simultaneously, relation modeling is performed between these features, efficiently obtaining the search region fusion features with better target-background discriminability for prediction. Furthermore, we introduce a novel feature selection mechanism based on modality reliability to mitigate the influence of invalid modalities for prediction, further improving the tracking performance. Extensive experiments on three popular RGB-T tracking benchmarks demonstrate that our method achieves new state-of-the-art performance while maintaining the fastest inference speed 84.2FPS. In particular, MPR/MSR on the short-term and long-term subsets of VTUAV dataset increased by 11.1$\%$/11.7$\%$ and 11.3$\%$/9.7$\%$.

Jiayu Zou, Zheng Zhu, Yun Ye et al.

BEV perception is of great importance in the field of autonomous driving, serving as the cornerstone of planning, controlling, and motion prediction. The quality of the BEV feature highly affects the performance of BEV perception. However, taking the noises in camera parameters and LiDAR scans into consideration, we usually obtain BEV representation with harmful noises. Diffusion models naturally have the ability to denoise noisy samples to the ideal data, which motivates us to utilize the diffusion model to get a better BEV representation. In this work, we propose an end-to-end framework, named DiffBEV, to exploit the potential of diffusion model to generate a more comprehensive BEV representation. To the best of our knowledge, we are the first to apply diffusion model to BEV perception. In practice, we design three types of conditions to guide the training of the diffusion model which denoises the coarse samples and refines the semantic feature in a progressive way. What's more, a cross-attention module is leveraged to fuse the context of BEV feature and the semantic content of conditional diffusion model. DiffBEV achieves a 25.9% mIoU on the nuScenes dataset, which is 6.2% higher than the best-performing existing approach. Quantitative and qualitative results on multiple benchmarks demonstrate the effectiveness of DiffBEV in BEV semantic segmentation and 3D object detection tasks. The code will be available soon.

Yanran Zhang, Ziyi Wang, Wenzhao Zheng et al.

Generating interactive and dynamic 4D scenes from a single static image remains a core challenge. Most existing generate-then-reconstruct and reconstruct-then-generate methods decouple geometry from motion, causing spatiotemporal inconsistencies and poor generalization. To address these, we extend the reconstruct-then-generate framework to jointly perform Motion generation and geometric Reconstruction for 4D Synthesis (MoRe4D). We first introduce TrajScene-60K, a large-scale dataset of 60,000 video samples with dense point trajectories, addressing the scarcity of high-quality 4D scene data. Based on this, we propose a diffusion-based 4D Scene Trajectory Generator (4D-STraG) to jointly generate geometrically consistent and motion-plausible 4D point trajectories. To leverage single-view priors, we design a depth-guided motion normalization strategy and a motion-aware module for effective geometry and dynamics integration. We then propose a 4D View Synthesis Module (4D-ViSM) to render videos with arbitrary camera trajectories from 4D point track representations. Experiments show that MoRe4D generates high-quality 4D scenes with multi-view consistency and rich dynamic details from a single image. Code: https://github.com/Zhangyr2022/MoRe4D.

Ming Wang, Beibei Lin, Xianda Guo et al.

Many gait recognition methods first partition the human gait into N-parts and then combine them to establish part-based feature representations. Their gait recognition performance is often affected by partitioning strategies, which are empirically chosen in different datasets. However, we observe that strips as the basic component of parts are agnostic against different partitioning strategies. Motivated by this observation, we present a strip-based multi-level gait recognition network, named GaitStrip, to extract comprehensive gait information at different levels. To be specific, our high-level branch explores the context of gait sequences and our low-level one focuses on detailed posture changes. We introduce a novel StriP-Based feature extractor (SPB) to learn the strip-based feature representations by directly taking each strip of the human body as the basic unit. Moreover, we propose a novel multi-branch structure, called Enhanced Convolution Module (ECM), to extract different representations of gaits. ECM consists of the Spatial-Temporal feature extractor (ST), the Frame-Level feature extractor (FL) and SPB, and has two obvious advantages: First, each branch focuses on a specific representation, which can be used to improve the robustness of the network. Specifically, ST aims to extract spatial-temporal features of gait sequences, while FL is used to generate the feature representation of each frame. Second, the parameters of the ECM can be reduced in test by introducing a structural re-parameterization technique. Extensive experimental results demonstrate that our GaitStrip achieves state-of-the-art performance in both normal walking and complex conditions.

Yiqun Duan, Xianda Guo, Zheng Zhu

Monocular depth estimation is a challenging task that predicts the pixel-wise depth from a single 2D image. Current methods typically model this problem as a regression or classification task. We propose DiffusionDepth, a new approach that reformulates monocular depth estimation as a denoising diffusion process. It learns an iterative denoising process to `denoise' random depth distribution into a depth map with the guidance of monocular visual conditions. The process is performed in the latent space encoded by a dedicated depth encoder and decoder. Instead of diffusing ground truth (GT) depth, the model learns to reverse the process of diffusing the refined depth of itself into random depth distribution. This self-diffusion formulation overcomes the difficulty of applying generative models to sparse GT depth scenarios. The proposed approach benefits this task by refining depth estimation step by step, which is superior for generating accurate and highly detailed depth maps. Experimental results on KITTI and NYU-Depth-V2 datasets suggest that a simple yet efficient diffusion approach could reach state-of-the-art performance in both indoor and outdoor scenarios with acceptable inference time.

Ming Wang, Xianda Guo, Beibei Lin et al.

Gait recognition is a biometric technology that recognizes the identity of humans through their walking patterns. Compared with other biometric technologies, gait recognition is more difficult to disguise and can be applied to the condition of long-distance without the cooperation of subjects. Thus, it has unique potential and wide application for crime prevention and social security. At present, most gait recognition methods directly extract features from the video frames to establish representations. However, these architectures learn representations from different features equally but do not pay enough attention to dynamic features, which refers to a representation of dynamic parts of silhouettes over time (e.g. legs). Since dynamic parts of the human body are more informative than other parts (e.g. bags) during walking, in this paper, we propose a novel and high-performance framework named DyGait. This is the first framework on gait recognition that is designed to focus on the extraction of dynamic features. Specifically, to take full advantage of the dynamic information, we propose a Dynamic Augmentation Module (DAM), which can automatically establish spatial-temporal feature representations of the dynamic parts of the human body. The experimental results show that our DyGait network outperforms other state-of-the-art gait recognition methods. It achieves an average Rank-1 accuracy of 71.4% on the GREW dataset, 66.3% on the Gait3D dataset, 98.4% on the CASIA-B dataset and 98.3% on the OU-MVLP dataset.

Jindi Lv, Yuhao Zhou, Zheng Zhu et al.

Chain-of-Thought (CoT) reasoning has significantly advanced the problem-solving capabilities of Large Language Models (LLMs), yet conventional CoT often exhibits internal determinism during decoding, limiting exploration of plausible alternatives. Recent methods attempt to address this by generating soft abstract tokens to enable reasoning in a continuous semantic space. However, we find that such approaches remain constrained by the greedy nature of autoregressive decoding, which fundamentally isolates the model from alternative reasoning possibilities. In this work, we propose Multi-Path Perception Policy Optimization (M3PO), a novel reinforcement learning framework that explicitly injects collective insights into the reasoning process. M3PO leverages parallel policy rollouts as naturally diverse reasoning sources and integrates cross-path interactions into policy updates through a lightweight collaborative mechanism. This design allows each trajectory to refine its reasoning with peer feedback, thereby cultivating more reliable multi-step reasoning patterns. Empirical results show that M3PO achieves state-of-the-art performance on both knowledge- and reasoning-intensive benchmarks. Models trained with M3PO maintain interpretability and inference efficiency, underscoring the promise of multi-path collaborative learning for robust reasoning.

Chaojun Ni, Cheng Chen, Xiaofeng Wang et al.

Vision-Language-Action (VLA) models built on pretrained Vision-Language Models (VLMs) show strong potential but are limited in practicality due to their large parameter counts. To mitigate this issue, using a lightweight VLM has been explored, but it compromises spatiotemporal reasoning. Although some methods suggest that incorporating additional 3D inputs can help, they usually rely on large VLMs to fuse 3D and 2D inputs and still lack temporal understanding. Therefore, we propose SwiftVLA, an architecture that enhances a compact model with 4D understanding while preserving design efficiency. Specifically, our approach features a pretrained 4D visual geometry transformer with a temporal cache that extracts 4D features from 2D images. Then, to enhance the VLM's ability to exploit both 2D images and 4D features, we introduce Fusion Tokens, a set of learnable tokens trained with a future prediction objective to generate unified representations for action generation. Finally, we introduce a mask-and-reconstruct strategy that masks 4D inputs to the VLM and trains the VLA to reconstruct them, enabling the VLM to learn effective 4D representations and allowing the 4D branch to be dropped at inference with minimal performance loss. Experiments in real and simulated environments show that SwiftVLA outperforms lightweight baselines and rivals VLAs up to 7 times larger, achieving comparable performance on edge devices while being 18 times faster and reducing memory footprint by 12 times.

Linkang Du, Zheng Zhu, Min Chen et al.

Text-to-image models based on diffusion processes, such as DALL-E, Stable Diffusion, and Midjourney, are capable of transforming texts into detailed images and have widespread applications in art and design. As such, amateur users can easily imitate professional-level paintings by collecting an artist's work and fine-tuning the model, leading to concerns about artworks' copyright infringement. To tackle these issues, previous studies either add visually imperceptible perturbation to the artwork to change its underlying styles (perturbation-based methods) or embed post-training detectable watermarks in the artwork (watermark-based methods). However, when the artwork or the model has been published online, i.e., modification to the original artwork or model retraining is not feasible, these strategies might not be viable. To this end, we propose a novel method for data-use auditing in the text-to-image generation model. The general idea of ArtistAuditor is to identify if a suspicious model has been finetuned using the artworks of specific artists by analyzing the features related to the style. Concretely, ArtistAuditor employs a style extractor to obtain the multi-granularity style representations and treats artworks as samplings of an artist's style. Then, ArtistAuditor queries a trained discriminator to gain the auditing decisions. The experimental results on six combinations of models and datasets show that ArtistAuditor can achieve high AUC values (> 0.937). By studying ArtistAuditor's transferability and core modules, we provide valuable insights into the practical implementation. Finally, we demonstrate the effectiveness of ArtistAuditor in real-world cases by an online platform Scenario. ArtistAuditor is open-sourced at https://github.com/Jozenn/ArtistAuditor.

Junjie Huang, Guan Huang, Zheng Zhu et al.

Autonomous driving perceives its surroundings for decision making, which is one of the most complex scenarios in visual perception. The success of paradigm innovation in solving the 2D object detection task inspires us to seek an elegant, feasible, and scalable paradigm for fundamentally pushing the performance boundary in this area. To this end, we contribute the BEVDet paradigm in this paper. BEVDet performs 3D object detection in Bird-Eye-View (BEV), where most target values are defined and route planning can be handily performed. We merely reuse existing modules to build its framework but substantially develop its performance by constructing an exclusive data augmentation strategy and upgrading the Non-Maximum Suppression strategy. In the experiment, BEVDet offers an excellent trade-off between accuracy and time-efficiency. As a fast version, BEVDet-Tiny scores 31.2% mAP and 39.2% NDS on the nuScenes val set. It is comparable with FCOS3D, but requires just 11% computational budget of 215.3 GFLOPs and runs 9.2 times faster at 15.6 FPS. Another high-precision version dubbed BEVDet-Base scores 39.3% mAP and 47.2% NDS, significantly exceeding all published results. With a comparable inference speed, it surpasses FCOS3D by a large margin of +9.8% mAP and +10.0% NDS. The source code is publicly available for further research at https://github.com/HuangJunJie2017/BEVDet .

Yongming Rao, Wenliang Zhao, Guangyi Chen et al.

Recent progress has shown that large-scale pre-training using contrastive image-text pairs can be a promising alternative for high-quality visual representation learning from natural language supervision. Benefiting from a broader source of supervision, this new paradigm exhibits impressive transferability to downstream classification tasks and datasets. However, the problem of transferring the knowledge learned from image-text pairs to more complex dense prediction tasks has barely been visited. In this work, we present a new framework for dense prediction by implicitly and explicitly leveraging the pre-trained knowledge from CLIP. Specifically, we convert the original image-text matching problem in CLIP to a pixel-text matching problem and use the pixel-text score maps to guide the learning of dense prediction models. By further using the contextual information from the image to prompt the language model, we are able to facilitate our model to better exploit the pre-trained knowledge. Our method is model-agnostic, which can be applied to arbitrary dense prediction systems and various pre-trained visual backbones including both CLIP models and ImageNet pre-trained models. Extensive experiments demonstrate the superior performance of our methods on semantic segmentation, object detection, and instance segmentation tasks. Code is available at https://github.com/raoyongming/DenseCLIP

Yunze Chen, Junjie Huang, Jiagang Zhu et al.

Recently, face recognition in the wild has achieved remarkable success and one key engine is the increasing size of training data. For example, the largest face dataset, WebFace42M contains about 2 million identities and 42 million faces. However, a massive number of faces raise the constraints in training time, computing resources, and memory cost. The current research on this problem mainly focuses on designing an efficient Fully-connected layer (FC) to reduce GPU memory consumption caused by a large number of identities. In this work, we relax these constraints by resolving the redundancy problem of the up-to-date face datasets caused by the greedily collecting operation (i.e. the core-set selection perspective). As the first attempt in this perspective on the face recognition problem, we find that existing methods are limited in both performance and efficiency. For superior cost-efficiency, we contribute a novel filtering strategy dubbed Face-NMS. Face-NMS works on feature space and simultaneously considers the local and global sparsity in generating core sets. In practice, Face-NMS is analogous to Non-Maximum Suppression (NMS) in the object detection community. It ranks the faces by their potential contribution to the overall sparsity and filters out the superfluous face in the pairs with high similarity for local sparsity. With respect to the efficiency aspect, Face-NMS accelerates the whole pipeline by applying a smaller but sufficient proxy dataset in training the proxy model. As a result, with Face-NMS, we successfully scale down the WebFace42M dataset to 60% while retaining its performance on the main benchmarks, offering a 40% resource-saving and 1.64 times acceleration. The code is publicly available for reference at https://github.com/HuangJunJie2017/Face-NMS.

Shuai Shen, Wanhua Li, Zheng Zhu et al.

Face clustering is a promising method for annotating unlabeled face images. Recent supervised approaches have boosted the face clustering accuracy greatly, however their performance is still far from satisfactory. These methods can be roughly divided into global-based and local-based ones. Global-based methods suffer from the limitation of training data scale, while local-based ones are difficult to grasp the whole graph structure information and usually take a long time for inference. Previous approaches fail to tackle these two challenges simultaneously. To address the dilemma of large-scale training and efficient inference, we propose the STructure-AwaRe Face Clustering (STAR-FC) method. Specifically, we design a structure-preserved subgraph sampling strategy to explore the power of large-scale training data, which can increase the training data scale from ${10^{5}}$ to ${10^{7}}$. During inference, the STAR-FC performs efficient full-graph clustering with two steps: graph parsing and graph refinement. And the concept of node intimacy is introduced in the second step to mine the local structural information. The STAR-FC gets 91.97 pairwise F-score on partial MS1M within 310s which surpasses the state-of-the-arts. Furthermore, we are the first to train on very large-scale graph with 20M nodes, and achieve superior inference results on 12M testing data. Overall, as a simple and effective method, the proposed STAR-FC provides a strong baseline for large-scale face clustering. Code is available at \url{https://sstzal.github.io/STAR-FC/}.

Junjie Huang, Zheng Zhu, Feng Guo et al.

Being a fundamental component in training and inference, data processing has not been systematically considered in human pose estimation community, to the best of our knowledge. In this paper, we focus on this problem and find that the devil of human pose estimation evolution is in the biased data processing. Specifically, by investigating the standard data processing in state-of-the-art approaches mainly including coordinate system transformation and keypoint format transformation (i.e., encoding and decoding), we find that the results obtained by common flipping strategy are unaligned with the original ones in inference. Moreover, there is a statistical error in some keypoint format transformation methods. Two problems couple together, significantly degrade the pose estimation performance and thus lay a trap for the research community. This trap has given bone to many suboptimal remedies, which are always unreported, confusing but influential. By causing failure in reproduction and unfair in comparison, the unreported remedies seriously impedes the technological development. To tackle this dilemma from the source, we propose Unbiased Data Processing (UDP) consist of two technique aspect for the two aforementioned problems respectively (i.e., unbiased coordinate system transformation and unbiased keypoint format transformation). As a model-agnostic approach and a superior solution, UDP successfully pushes the performance boundary of human pose estimation and offers a higher and more reliable baseline for research community. Code is public available in https://github.com/HuangJunJie2017/UDP-Pose

Guosheng Zhao, Chaojun Ni, Xiaofeng Wang et al.

Closed-loop simulation is essential for advancing end-to-end autonomous driving systems. Contemporary sensor simulation methods, such as NeRF and 3DGS, rely predominantly on conditions closely aligned with training data distributions, which are largely confined to forward-driving scenarios. Consequently, these methods face limitations when rendering complex maneuvers (e.g., lane change, acceleration, deceleration). Recent advancements in autonomous-driving world models have demonstrated the potential to generate diverse driving videos. However, these approaches remain constrained to 2D video generation, inherently lacking the spatiotemporal coherence required to capture intricacies of dynamic driving environments. In this paper, we introduce DriveDreamer4D, which enhances 4D driving scene representation leveraging world model priors. Specifically, we utilize the world model as a data machine to synthesize novel trajectory videos, where structured conditions are explicitly leveraged to control the spatial-temporal consistency of traffic elements. Besides, the cousin data training strategy is proposed to facilitate merging real and synthetic data for optimizing 4DGS. To our knowledge, DriveDreamer4D is the first to utilize video generation models for improving 4D reconstruction in driving scenarios. Experimental results reveal that DriveDreamer4D significantly enhances generation quality under novel trajectory views, achieving a relative improvement in FID by 32.1%, 46.4%, and 16.3% compared to PVG, S3Gaussian, and Deformable-GS. Moreover, DriveDreamer4D markedly enhances the spatiotemporal coherence of driving agents, which is verified by a comprehensive user study and the relative increases of 22.6%, 43.5%, and 15.6% in the NTA-IoU metric.

Chen Min, Dawei Zhao, Liang Xiao et al.

Vision-centric autonomous driving has recently raised wide attention due to its lower cost. Pre-training is essential for extracting a universal representation. However, current vision-centric pre-training typically relies on either 2D or 3D pre-text tasks, overlooking the temporal characteristics of autonomous driving as a 4D scene understanding task. In this paper, we address this challenge by introducing a world model-based autonomous driving 4D representation learning framework, dubbed \emph{DriveWorld}, which is capable of pre-training from multi-camera driving videos in a spatio-temporal fashion. Specifically, we propose a Memory State-Space Model for spatio-temporal modelling, which consists of a Dynamic Memory Bank module for learning temporal-aware latent dynamics to predict future changes and a Static Scene Propagation module for learning spatial-aware latent statics to offer comprehensive scene contexts. We additionally introduce a Task Prompt to decouple task-aware features for various downstream tasks. The experiments demonstrate that DriveWorld delivers promising results on various autonomous driving tasks. When pre-trained with the OpenScene dataset, DriveWorld achieves a 7.5% increase in mAP for 3D object detection, a 3.0% increase in IoU for online mapping, a 5.0% increase in AMOTA for multi-object tracking, a 0.1m decrease in minADE for motion forecasting, a 3.0% increase in IoU for occupancy prediction, and a 0.34m reduction in average L2 error for planning.

Chaojun Ni, Guosheng Zhao, Xiaofeng Wang et al.

Closed-loop simulation is crucial for end-to-end autonomous driving. Existing sensor simulation methods (e.g., NeRF and 3DGS) reconstruct driving scenes based on conditions that closely mirror training data distributions. However, these methods struggle with rendering novel trajectories, such as lane changes. Recent works have demonstrated that integrating world model knowledge alleviates these issues. Despite their efficiency, these approaches still encounter difficulties in the accurate representation of more complex maneuvers, with multi-lane shifts being a notable example. Therefore, we introduce ReconDreamer, which enhances driving scene reconstruction through incremental integration of world model knowledge. Specifically, DriveRestorer is proposed to mitigate artifacts via online restoration. This is complemented by a progressive data update strategy designed to ensure high-quality rendering for more complex maneuvers. To the best of our knowledge, ReconDreamer is the first method to effectively render in large maneuvers. Experimental results demonstrate that ReconDreamer outperforms Street Gaussians in the NTA-IoU, NTL-IoU, and FID, with relative improvements by 24.87%, 6.72%, and 29.97%. Furthermore, ReconDreamer surpasses DriveDreamer4D with PVG during large maneuver rendering, as verified by a relative improvement of 195.87% in the NTA-IoU metric and a comprehensive user study.

Xiaofeng Wang, Kang Zhao, Feng Liu et al.

Video generation has emerged as a promising tool for world simulation, leveraging visual data to replicate real-world environments. Within this context, egocentric video generation, which centers on the human perspective, holds significant potential for enhancing applications in virtual reality, augmented reality, and gaming. However, the generation of egocentric videos presents substantial challenges due to the dynamic nature of egocentric viewpoints, the intricate diversity of actions, and the complex variety of scenes encountered. Existing datasets are inadequate for addressing these challenges effectively. To bridge this gap, we present EgoVid-5M, the first high-quality dataset specifically curated for egocentric video generation. EgoVid-5M encompasses 5 million egocentric video clips and is enriched with detailed action annotations, including fine-grained kinematic control and high-level textual descriptions. To ensure the integrity and usability of the dataset, we implement a sophisticated data cleaning pipeline designed to maintain frame consistency, action coherence, and motion smoothness under egocentric conditions. Furthermore, we introduce EgoDreamer, which is capable of generating egocentric videos driven simultaneously by action descriptions and kinematic control signals. The EgoVid-5M dataset, associated action annotations, and all data cleansing metadata will be released for the advancement of research in egocentric video generation.

Boyuan Wang, Xiaofeng Wang, Chaojun Ni et al.

Human-motion video generation has been a challenging task, primarily due to the difficulty inherent in learning human body movements. While some approaches have attempted to drive human-centric video generation explicitly through pose control, these methods typically rely on poses derived from existing videos, thereby lacking flexibility. To address this, we propose HumanDreamer, a decoupled human video generation framework that first generates diverse poses from text prompts and then leverages these poses to generate human-motion videos. Specifically, we propose MotionVid, the largest dataset for human-motion pose generation. Based on the dataset, we present MotionDiT, which is trained to generate structured human-motion poses from text prompts. Besides, a novel LAMA loss is introduced, which together contribute to a significant improvement in FID by 62.4%, along with respective enhancements in R-precision for top1, top2, and top3 by 41.8%, 26.3%, and 18.3%, thereby advancing both the Text-to-Pose control accuracy and FID metrics. Our experiments across various Pose-to-Video baselines demonstrate that the poses generated by our method can produce diverse and high-quality human-motion videos. Furthermore, our model can facilitate other downstream tasks, such as pose sequence prediction and 2D-3D motion lifting.

Boyuan Wang, Runqi Ouyang, Xiaofeng Wang et al.

Single-image human reconstruction is vital for digital human modeling applications but remains an extremely challenging task. Current approaches rely on generative models to synthesize multi-view images for subsequent 3D reconstruction and animation. However, directly generating multiple views from a single human image suffers from geometric inconsistencies, resulting in issues like fragmented or blurred limbs in the reconstructed models. To tackle these limitations, we introduce \textbf{HumanDreamer-X}, a novel framework that integrates multi-view human generation and reconstruction into a unified pipeline, which significantly enhances the geometric consistency and visual fidelity of the reconstructed 3D models. In this framework, 3D Gaussian Splatting serves as an explicit 3D representation to provide initial geometry and appearance priority. Building upon this foundation, \textbf{HumanFixer} is trained to restore 3DGS renderings, which guarantee photorealistic results. Furthermore, we delve into the inherent challenges associated with attention mechanisms in multi-view human generation, and propose an attention modulation strategy that effectively enhances geometric details identity consistency across multi-view. Experimental results demonstrate that our approach markedly improves generation and reconstruction PSNR quality metrics by 16.45% and 12.65%, respectively, achieving a PSNR of up to 25.62 dB, while also showing generalization capabilities on in-the-wild data and applicability to various human reconstruction backbone models.

Chaojun Ni, Guosheng Zhao, Xiaofeng Wang et al.

Reinforcement learning for training end-to-end autonomous driving models in closed-loop simulations is gaining growing attention. However, most simulation environments differ significantly from real-world conditions, creating a substantial simulation-to-reality (sim2real) gap. To bridge this gap, some approaches utilize scene reconstruction techniques to create photorealistic environments as a simulator. While this improves realistic sensor simulation, these methods are inherently constrained by the distribution of the training data, making it difficult to render high-quality sensor data for novel trajectories or corner case scenarios. Therefore, we propose ReconDreamer-RL, a framework designed to integrate video diffusion priors into scene reconstruction to aid reinforcement learning, thereby enhancing end-to-end autonomous driving training. Specifically, in ReconDreamer-RL, we introduce ReconSimulator, which combines the video diffusion prior for appearance modeling and incorporates a kinematic model for physical modeling, thereby reconstructing driving scenarios from real-world data. This narrows the sim2real gap for closed-loop evaluation and reinforcement learning. To cover more corner-case scenarios, we introduce the Dynamic Adversary Agent (DAA), which adjusts the trajectories of surrounding vehicles relative to the ego vehicle, autonomously generating corner-case traffic scenarios (e.g., cut-in). Finally, the Cousin Trajectory Generator (CTG) is proposed to address the issue of training data distribution, which is often biased toward simple straight-line movements. Experiments show that ReconDreamer-RL improves end-to-end autonomous driving training, outperforming imitation learning methods with a 5x reduction in the Collision Ratio.

Zhehao Dong, Xiaofeng Wang, Zheng Zhu et al.

Vision-language-action (VLA) models increasingly rely on diverse training data to achieve robust generalization. However, collecting large-scale real-world robot manipulation data across varied object appearances and environmental conditions remains prohibitively time-consuming and expensive. To overcome this bottleneck, we propose Embodied Manipulation Media Adaptation (EMMA), a VLA policy enhancement framework that integrates a generative data engine with an effective training pipeline. We introduce DreamTransfer, a diffusion Transformer-based framework for generating multi-view consistent, geometrically grounded embodied manipulation videos. DreamTransfer enables text-controlled visual editing of robot videos, transforming foreground, background, and lighting conditions without compromising 3D structure or geometrical plausibility. Furthermore, we explore hybrid training with real and generated data, and introduce AdaMix, a hard-sample-aware training strategy that dynamically reweights training batches to focus optimization on perceptually or kinematically challenging samples. Extensive experiments show that videos generated by DreamTransfer significantly outperform prior video generation methods in multi-view consistency, geometric fidelity, and text-conditioning accuracy. Crucially, VLAs trained with generated data enable robots to generalize to unseen object categories and novel visual domains using only demonstrations from a single appearance. In real-world robotic manipulation tasks with zero-shot visual domains, our approach achieves over a 200% relative performance gain compared to training on real data alone, and further improves by 13% with AdaMix, demonstrating its effectiveness in boosting policy generalization.

Chaojun Ni, Jie Li, Haoyun Li et al.

Interactive 3D scene generation from a single image has gained significant attention due to its potential to create immersive virtual worlds. However, a key challenge in current 3D generation methods is the limited explorability, which cannot render high-quality images during larger maneuvers beyond the original viewpoint, particularly when attempting to move forward into unseen areas. To address this challenge, we propose WonderFree, the first model that enables users to interactively generate 3D worlds with the freedom to explore from arbitrary angles and directions. Specifically, we decouple this challenge into two key subproblems: novel view quality, which addresses visual artifacts and floating issues in novel views, and cross-view consistency, which ensures spatial consistency across different viewpoints. To enhance rendering quality in novel views, we introduce WorldRestorer, a data-driven video restoration model designed to eliminate floaters and artifacts. In addition, a data collection pipeline is presented to automatically gather training data for WorldRestorer, ensuring it can handle scenes with varying styles needed for 3D scene generation. Furthermore, to improve cross-view consistency, we propose ConsistView, a multi-view joint restoration mechanism that simultaneously restores multiple perspectives while maintaining spatiotemporal coherence. Experimental results demonstrate that WonderFree not only enhances rendering quality across diverse viewpoints but also significantly improves global coherence and consistency. These improvements are confirmed by CLIP-based metrics and a user study showing a 77.20% preference for WonderFree over WonderWorld enabling a seamless and immersive 3D exploration experience. The code, model, and data will be publicly available.

Weijie Wang, Yeqing Chen, Zeyu Zhang et al.

Feed-forward 3D Gaussian Splatting (3DGS) has emerged as a highly effective solution for novel view synthesis. Existing methods predominantly rely on a pixel-aligned Gaussian prediction paradigm, where each 2D pixel is mapped to a 3D Gaussian. We rethink this widely adopted formulation and identify several inherent limitations: it renders the reconstructed 3D models heavily dependent on the number of input views, leads to view-biased density distributions, and introduces alignment errors, particularly when source views contain occlusions or low texture. To address these challenges, we introduce VolSplat, a new multi-view feed-forward paradigm that replaces pixel alignment with voxel-aligned Gaussians. By directly predicting Gaussians from a predicted 3D voxel grid, it overcomes pixel alignment's reliance on error-prone 2D feature matching, ensuring robust multi-view consistency. Furthermore, it enables adaptive control over Gaussian density based on 3D scene complexity, yielding more faithful Gaussian point clouds, improved geometric consistency, and enhanced novel-view rendering quality. Experiments on widely used benchmarks including RealEstate10K and ScanNet demonstrate that VolSplat achieves state-of-the-art performance while producing more plausible and view-consistent Gaussian reconstructions. In addition to superior results, our approach establishes a more scalable framework for feed-forward 3D reconstruction with denser and more robust representations, paving the way for further research in wider communities. The video results, code and trained models are available on our project page: https://lhmd.top/volsplat.

Yifan Chang, Jie Qin, Limeng Qiao et al.

Vector quantization (VQ) is a key component in discrete tokenizers for image generation, but its training is often unstable due to straight-through estimation bias, one-step-behind updates, and sparse codebook gradients, which lead to suboptimal reconstruction performance and low codebook usage. In this work, we analyze these fundamental challenges and provide a simple yet effective solution. To maintain high codebook usage in VQ networks (VQN) during learning annealing and codebook size expansion, we propose VQBridge, a robust, scalable, and efficient projector based on the map function method. VQBridge optimizes code vectors through a compress-process-recover pipeline, enabling stable and effective codebook training. By combining VQBridge with learning annealing, our VQN achieves full (100%) codebook usage across diverse codebook configurations, which we refer to as FVQ (FullVQ). Through extensive experiments, we demonstrate that FVQ is effective, scalable, and generalizable: it attains 100% codebook usage even with a 262k-codebook, achieves state-of-the-art reconstruction performance, consistently improves with larger codebooks, higher vector channels, or longer training, and remains effective across different VQ variants. Moreover, when integrated with LlamaGen, FVQ significantly enhances image generation performance, surpassing visual autoregressive models (VAR) by 0.5 and diffusion models (DiT) by 0.2 rFID, highlighting the importance of high-quality tokenizers for strong autoregressive image generation.

Penghao Wang, Yuhao Zhou, Mengxuan Wu et al.

As large language models (LLMs) advance, the ultimate vision for their role in science is emerging: we could build an AI collaborator to effectively assist human beings throughout the entire scientific research process. We refer to this envisioned system as ResearchGPT. Given that scientific research progresses through multiple interdependent phases, achieving this vision requires rigorous benchmarks that evaluate the end-to-end workflow rather than isolated sub-tasks. To this end, we contribute CS-54k, a high-quality corpus of scientific Q&A pairs in computer science, built from 14k CC-licensed papers. It is constructed through a scalable, paper-grounded pipeline that combines retrieval-augmented generation (RAG) with multi-stage quality control to ensure factual grounding. From this unified corpus, we derive two complementary subsets: CS-4k, a carefully curated benchmark for evaluating AI's ability to assist scientific research, and CS-50k, a large-scale training dataset. Extensive experiments demonstrate that CS-4k stratifies state-of-the-art LLMs into distinct capability tiers. Open models trained on CS-50k with supervised training and reinforcement learning demonstrate substantial improvements. Even 7B-scale models, when properly trained, outperform many larger proprietary systems, such as GPT-4.1, GPT-4o, and Gemini 2.5 Pro. This indicates that making AI models better research assistants relies more on domain-aligned training with high-quality data than on pretraining scale or general benchmark performance. We release CS-4k and CS-50k in the hope of fostering AI systems as reliable collaborators in CS research.