Jiwen Lu

Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang et al. · tsinghua

Modern methods for vision-centric autonomous driving perception widely adopt the bird's-eye-view (BEV) representation to describe a 3D scene. Despite its better efficiency than voxel representation, it has difficulty describing the fine-grained 3D structure of a scene with a single plane. To address this, we propose a tri-perspective view (TPV) representation which accompanies BEV with two additional perpendicular planes. We model each point in the 3D space by summing its projected features on the three planes. To lift image features to the 3D TPV space, we further propose a transformer-based TPV encoder (TPVFormer) to obtain the TPV features effectively. We employ the attention mechanism to aggregate the image features corresponding to each query in each TPV plane. Experiments show that our model trained with sparse supervision effectively predicts the semantic occupancy for all voxels. We demonstrate for the first time that using only camera inputs can achieve comparable performance with LiDAR-based methods on the LiDAR segmentation task on nuScenes. Code: https://github.com/wzzheng/TPVFormer.

Wenliang Zhao, Lujia Bai, Yongming Rao et al. · tsinghua

Diffusion probabilistic models (DPMs) have demonstrated a very promising ability in high-resolution image synthesis. However, sampling from a pre-trained DPM is time-consuming due to the multiple evaluations of the denoising network, making it more and more important to accelerate the sampling of DPMs. Despite recent progress in designing fast samplers, existing methods still cannot generate satisfying images in many applications where fewer steps (e.g., $<$10) are favored. In this paper, we develop a unified corrector (UniC) that can be applied after any existing DPM sampler to increase the order of accuracy without extra model evaluations, and derive a unified predictor (UniP) that supports arbitrary order as a byproduct. Combining UniP and UniC, we propose a unified predictor-corrector framework called UniPC for the fast sampling of DPMs, which has a unified analytical form for any order and can significantly improve the sampling quality over previous methods, especially in extremely few steps. We evaluate our methods through extensive experiments including both unconditional and conditional sampling using pixel-space and latent-space DPMs. Our UniPC can achieve 3.87 FID on CIFAR10 (unconditional) and 7.51 FID on ImageNet 256$\times$256 (conditional) with only 10 function evaluations. Code is available at https://github.com/wl-zhao/UniPC.

Yongming Rao, Wenliang Zhao, Yansong Tang et al. · tsinghua

Recent progress in vision Transformers exhibits great success in various tasks driven by the new spatial modeling mechanism based on dot-product self-attention. In this paper, we show that the key ingredients behind the vision Transformers, namely input-adaptive, long-range and high-order spatial interactions, can also be efficiently implemented with a convolution-based framework. We present the Recursive Gated Convolution ($\textit{g}^\textit{n}$Conv) that performs high-order spatial interactions with gated convolutions and recursive designs. The new operation is highly flexible and customizable, which is compatible with various variants of convolution and extends the two-order interactions in self-attention to arbitrary orders without introducing significant extra computation. $\textit{g}^\textit{n}$Conv can serve as a plug-and-play module to improve various vision Transformers and convolution-based models. Based on the operation, we construct a new family of generic vision backbones named HorNet. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation show HorNet outperform Swin Transformers and ConvNeXt by a significant margin with similar overall architecture and training configurations. HorNet also shows favorable scalability to more training data and larger model sizes. Apart from the effectiveness in visual encoders, we also show $\textit{g}^\textit{n}$Conv can be applied to task-specific decoders and consistently improve dense prediction performance with less computation. Our results demonstrate that $\textit{g}^\textit{n}$Conv can be a new basic module for visual modeling that effectively combines the merits of both vision Transformers and CNNs. Code is available at https://github.com/raoyongming/HorNet

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

Tianpei Gu, Guangyi Chen, Junlong Li et al. · tsinghua

Human behavior has the nature of indeterminacy, which requires the pedestrian trajectory prediction system to model the multi-modality of future motion states. Unlike existing stochastic trajectory prediction methods which usually use a latent variable to represent multi-modality, we explicitly simulate the process of human motion variation from indeterminate to determinate. In this paper, we present a new framework to formulate the trajectory prediction task as a reverse process of motion indeterminacy diffusion (MID), in which we progressively discard indeterminacy from all the walkable areas until reaching the desired trajectory. This process is learned with a parameterized Markov chain conditioned by the observed trajectories. We can adjust the length of the chain to control the degree of indeterminacy and balance the diversity and determinacy of the predictions. Specifically, we encode the history behavior information and the social interactions as a state embedding and devise a Transformer-based diffusion model to capture the temporal dependencies of trajectories. Extensive experiments on the human trajectory prediction benchmarks including the Stanford Drone and ETH/UCY datasets demonstrate the superiority of our method. Code is available at https://github.com/gutianpei/MID.

Wenliang Zhao, Yongming Rao, Zuyan Liu et al. · tsinghua

Diffusion models (DMs) have become the new trend of generative models and have demonstrated a powerful ability of conditional synthesis. Among those, text-to-image diffusion models pre-trained on large-scale image-text pairs are highly controllable by customizable prompts. Unlike the unconditional generative models that focus on low-level attributes and details, text-to-image diffusion models contain more high-level knowledge thanks to the vision-language pre-training. In this paper, we propose VPD (Visual Perception with a pre-trained Diffusion model), a new framework that exploits the semantic information of a pre-trained text-to-image diffusion model in visual perception tasks. Instead of using the pre-trained denoising autoencoder in a diffusion-based pipeline, we simply use it as a backbone and aim to study how to take full advantage of the learned knowledge. Specifically, we prompt the denoising decoder with proper textual inputs and refine the text features with an adapter, leading to a better alignment to the pre-trained stage and making the visual contents interact with the text prompts. We also propose to utilize the cross-attention maps between the visual features and the text features to provide explicit guidance. Compared with other pre-training methods, we show that vision-language pre-trained diffusion models can be faster adapted to downstream visual perception tasks using the proposed VPD. Extensive experiments on semantic segmentation, referring image segmentation and depth estimation demonstrates the effectiveness of our method. Notably, VPD attains 0.254 RMSE on NYUv2 depth estimation and 73.3% oIoU on RefCOCO-val referring image segmentation, establishing new records on these two benchmarks. Code is available at https://github.com/wl-zhao/VPD

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.

Yifei Li, Pengyiang Liu, Yuhang Zang et al.

Multimodal agents in robotics, AR, and autonomous driving must reason about places and layouts from continuous egocentric streams, often using evidence outside the current view. Existing benchmarks either evaluate offline over full videos or target events rather than spatial structure. We introduce OVO-S-Bench, a fully human-annotated benchmark for streaming spatial intelligence, comprising 1,680 questions over 348 source videos. Annotation involves 12 trained annotators, each also serving as a blind cross-reviewer, across roughly 804 person-hours of multi-round quality assurance. Each question carries a query timestamp and an evidence interval, and at evaluation, the model sees only the prefix preceding the query. Questions span four levels of increasing abstraction: instantaneous egocentric perception, spatiotemporal context tracking, spatial simulation and reasoning, and allocentric mapping. Across 38 proprietary and open-source MLLMs, Gemini-3.1-Pro trails human experts by 27 points, 59.2 vs. 86.6, with allocentric mapping as the dominant bottleneck. Notably, streaming and spatially fine-tuned MLLMs underperform their own backbones. We further find that chain-of-thought reasoning amplifies spatial errors when ungrounded in the stream. By exposing these limitations, OVO-S-Bench establishes a demanding testbed for next-generation streaming spatial MLLMs.

Muheng Li, Yueqi Duan, Jie Zhou et al. · tsinghua

With the rising industrial attention to 3D virtual modeling technology, generating novel 3D content based on specified conditions (e.g. text) has become a hot issue. In this paper, we propose a new generative 3D modeling framework called Diffusion-SDF for the challenging task of text-to-shape synthesis. Previous approaches lack flexibility in both 3D data representation and shape generation, thereby failing to generate highly diversified 3D shapes conforming to the given text descriptions. To address this, we propose a SDF autoencoder together with the Voxelized Diffusion model to learn and generate representations for voxelized signed distance fields (SDFs) of 3D shapes. Specifically, we design a novel UinU-Net architecture that implants a local-focused inner network inside the standard U-Net architecture, which enables better reconstruction of patch-independent SDF representations. We extend our approach to further text-to-shape tasks including text-conditioned shape completion and manipulation. Experimental results show that Diffusion-SDF generates both higher quality and more diversified 3D shapes that conform well to given text descriptions when compared to previous approaches. Code is available at: https://github.com/ttlmh/Diffusion-SDF

Jinglin Xu, Yongming Rao, Xumin Yu et al. · tsinghua

Most existing action quality assessment methods rely on the deep features of an entire video to predict the score, which is less reliable due to the non-transparent inference process and poor interpretability. We argue that understanding both high-level semantics and internal temporal structures of actions in competitive sports videos is the key to making predictions accurate and interpretable. Towards this goal, we construct a new fine-grained dataset, called FineDiving, developed on diverse diving events with detailed annotations on action procedures. We also propose a procedure-aware approach for action quality assessment, learned by a new Temporal Segmentation Attention module. Specifically, we propose to parse pairwise query and exemplar action instances into consecutive steps with diverse semantic and temporal correspondences. The procedure-aware cross-attention is proposed to learn embeddings between query and exemplar steps to discover their semantic, spatial, and temporal correspondences, and further serve for fine-grained contrastive regression to derive a reliable scoring mechanism. Extensive experiments demonstrate that our approach achieves substantial improvements over state-of-the-art methods with better interpretability. The dataset and code are available at \url{https://github.com/xujinglin/FineDiving}.

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/.

Yuanhui Huang, Wenzhao Zheng, Borui Zhang et al. · tsinghua

3D occupancy prediction is an important task for the robustness of vision-centric autonomous driving, which aims to predict whether each point is occupied in the surrounding 3D space. Existing methods usually require 3D occupancy labels to produce meaningful results. However, it is very laborious to annotate the occupancy status of each voxel. In this paper, we propose SelfOcc to explore a self-supervised way to learn 3D occupancy using only video sequences. We first transform the images into the 3D space (e.g., bird's eye view) to obtain 3D representation of the scene. We directly impose constraints on the 3D representations by treating them as signed distance fields. We can then render 2D images of previous and future frames as self-supervision signals to learn the 3D representations. We propose an MVS-embedded strategy to directly optimize the SDF-induced weights with multiple depth proposals. Our SelfOcc outperforms the previous best method SceneRF by 58.7% using a single frame as input on SemanticKITTI and is the first self-supervised work that produces reasonable 3D occupancy for surround cameras on nuScenes. SelfOcc produces high-quality depth and achieves state-of-the-art results on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI, KITTI-2015, and nuScenes, respectively. Code: https://github.com/huang-yh/SelfOcc.

Xumin Yu, Yongming Rao, Ziyi Wang et al. · tsinghua

In this paper, we present a new method that reformulates point cloud completion as a set-to-set translation problem and design a new model, called PoinTr, which adopts a Transformer encoder-decoder architecture for point cloud completion. By representing the point cloud as a set of unordered groups of points with position embeddings, we convert the input data to a sequence of point proxies and employ the Transformers for generation. To facilitate Transformers to better leverage the inductive bias about 3D geometric structures of point clouds, we further devise a geometry-aware block that models the local geometric relationships explicitly. The migration of Transformers enables our model to better learn structural knowledge and preserve detailed information for point cloud completion. Taking a step towards more complicated and diverse situations, we further propose AdaPoinTr by developing an adaptive query generation mechanism and designing a novel denoising task during completing a point cloud. Coupling these two techniques enables us to train the model efficiently and effectively: we reduce training time (by 15x or more) and improve completion performance (over 20%). We also show our method can be extended to the scene-level point cloud completion scenario by designing a new geometry-enhanced semantic scene completion framework. Extensive experiments on the existing and newly-proposed datasets demonstrate the effectiveness of our method, which attains 6.53 CD on PCN, 0.81 CD on ShapeNet-55 and 0.392 MMD on real-world KITTI, surpassing other work by a large margin and establishing new state-of-the-arts on various benchmarks. Most notably, AdaPoinTr can achieve such promising performance with higher throughputs and fewer FLOPs compared with the previous best methods in practice. The code and datasets are available at https://github.com/yuxumin/PoinTr

Ziyi Wang, Xumin Yu, Yongming Rao et al. · tsinghua

Nowadays, pre-training big models on large-scale datasets has become a crucial topic in deep learning. The pre-trained models with high representation ability and transferability achieve a great success and dominate many downstream tasks in natural language processing and 2D vision. However, it is non-trivial to promote such a pretraining-tuning paradigm to the 3D vision, given the limited training data that are relatively inconvenient to collect. In this paper, we provide a new perspective of leveraging pre-trained 2D knowledge in 3D domain to tackle this problem, tuning pre-trained image models with the novel Point-to-Pixel prompting for point cloud analysis at a minor parameter cost. Following the principle of prompting engineering, we transform point clouds into colorful images with geometry-preserved projection and geometry-aware coloring to adapt to pre-trained image models, whose weights are kept frozen during the end-to-end optimization of point cloud analysis tasks. We conduct extensive experiments to demonstrate that cooperating with our proposed Point-to-Pixel Prompting, better pre-trained image model will lead to consistently better performance in 3D vision. Enjoying prosperous development from image pre-training field, our method attains 89.3% accuracy on the hardest setting of ScanObjectNN, surpassing conventional point cloud models with much fewer trainable parameters. Our framework also exhibits very competitive performance on ModelNet classification and ShapeNet Part Segmentation. Code is available at https://github.com/wangzy22/P2P.

Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang et al. · tsinghua

Semantic segmentation in autonomous driving has been undergoing an evolution from sparse point segmentation to dense voxel segmentation, where the objective is to predict the semantic occupancy of each voxel in the concerned 3D space. The dense nature of the prediction space has rendered existing efficient 2D-projection-based methods (e.g., bird's eye view, range view, etc.) ineffective, as they can only describe a subspace of the 3D scene. To address this, we propose a cylindrical tri-perspective view to represent point clouds effectively and comprehensively and a PointOcc model to process them efficiently. Considering the distance distribution of LiDAR point clouds, we construct the tri-perspective view in the cylindrical coordinate system for more fine-grained modeling of nearer areas. We employ spatial group pooling to maintain structural details during projection and adopt 2D backbones to efficiently process each TPV plane. Finally, we obtain the features of each point by aggregating its projected features on each of the processed TPV planes without the need for any post-processing. Extensive experiments on both 3D occupancy prediction and LiDAR segmentation benchmarks demonstrate that the proposed PointOcc achieves state-of-the-art performance with much faster speed. Specifically, despite only using LiDAR, PointOcc significantly outperforms all other methods, including multi-modal methods, with a large margin on the OpenOccupancy benchmark. Code: https://github.com/wzzheng/PointOcc.

Han Xiao, Ziwei Wang, Zheng Zhu et al. · tsinghua

In this paper, we propose a Shapley value based method to evaluate operation contribution (Shapley-NAS) for neural architecture search. Differentiable architecture search (DARTS) acquires the optimal architectures by optimizing the architecture parameters with gradient descent, which significantly reduces the search cost. However, the magnitude of architecture parameters updated by gradient descent fails to reveal the actual operation importance to the task performance and therefore harms the effectiveness of obtained architectures. By contrast, we propose to evaluate the direct influence of operations on validation accuracy. To deal with the complex relationships between supernet components, we leverage Shapley value to quantify their marginal contributions by considering all possible combinations. Specifically, we iteratively optimize the supernet weights and update the architecture parameters by evaluating operation contributions via Shapley value, so that the optimal architectures are derived by selecting the operations that contribute significantly to the tasks. Since the exact computation of Shapley value is NP-hard, the Monte-Carlo sampling based algorithm with early truncation is employed for efficient approximation, and the momentum update mechanism is adopted to alleviate fluctuation of the sampling process. Extensive experiments on various datasets and various search spaces show that our Shapley-NAS outperforms the state-of-the-art methods by a considerable margin with light search cost. The code is available at https://github.com/Euphoria16/Shapley-NAS.git

Yongming Rao, Zuyan Liu, Wenliang Zhao et al. · tsinghua

In this paper, we present a new approach for model acceleration by exploiting spatial sparsity in visual data. We observe that the final prediction in vision Transformers is only based on a subset of the most informative tokens, which is sufficient for accurate image recognition. Based on this observation, we propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically based on the input to accelerate vision Transformers. Specifically, we devise a lightweight prediction module to estimate the importance score of each token given the current features. The module is added to different layers to prune redundant tokens hierarchically. While the framework is inspired by our observation of the sparse attention in vision Transformers, we find the idea of adaptive and asymmetric computation can be a general solution for accelerating various architectures. We extend our method to hierarchical models including CNNs and hierarchical vision Transformers as well as more complex dense prediction tasks that require structured feature maps by formulating a more generic dynamic spatial sparsification framework with progressive sparsification and asymmetric computation for different spatial locations. By applying lightweight fast paths to less informative features and using more expressive slow paths to more important locations, we can maintain the structure of feature maps while significantly reducing the overall computations. Extensive experiments demonstrate the effectiveness of our framework on various modern architectures and different visual recognition tasks. Our results clearly demonstrate that dynamic spatial sparsification offers a new and more effective dimension for model acceleration. Code is available at https://github.com/raoyongming/DynamicViT

Wanhua Li, Xiaoke Huang, Zheng Zhu et al. · tsinghua

This paper presents a language-powered paradigm for ordinal regression. Existing methods usually treat each rank as a category and employ a set of weights to learn these concepts. These methods are easy to overfit and usually attain unsatisfactory performance as the learned concepts are mainly derived from the training set. Recent large pre-trained vision-language models like CLIP have shown impressive performance on various visual tasks. In this paper, we propose to learn the rank concepts from the rich semantic CLIP latent space. Specifically, we reformulate this task as an image-language matching problem with a contrastive objective, which regards labels as text and obtains a language prototype from a text encoder for each rank. While prompt engineering for CLIP is extremely time-consuming, we propose OrdinalCLIP, a differentiable prompting method for adapting CLIP for ordinal regression. OrdinalCLIP consists of learnable context tokens and learnable rank embeddings; The learnable rank embeddings are constructed by explicitly modeling numerical continuity, resulting in well-ordered, compact language prototypes in the CLIP space. Once learned, we can only save the language prototypes and discard the huge language model, resulting in zero additional computational overhead compared with the linear head counterpart. Experimental results show that our paradigm achieves competitive performance in general ordinal regression tasks, and gains improvements in few-shot and distribution shift settings for age estimation. The code is available at https://github.com/xk-huang/OrdinalCLIP.

Wanhua Li, Zhexuan Cao, Jianjiang Feng et al. · tsinghua

Objects are usually associated with multiple attributes, and these attributes often exhibit high correlations. Modeling complex relationships between attributes poses a great challenge for multi-attribute learning. This paper proposes a simple yet generic framework named Label2Label to exploit the complex attribute correlations. Label2Label is the first attempt for multi-attribute prediction from the perspective of language modeling. Specifically, it treats each attribute label as a "word" describing the sample. As each sample is annotated with multiple attribute labels, these "words" will naturally form an unordered but meaningful "sentence", which depicts the semantic information of the corresponding sample. Inspired by the remarkable success of pre-training language models in NLP, Label2Label introduces an image-conditioned masked language model, which randomly masks some of the "word" tokens from the label "sentence" and aims to recover them based on the masked "sentence" and the context conveyed by image features. Our intuition is that the instance-wise attribute relations are well grasped if the neural net can infer the missing attributes based on the context and the remaining attribute hints. Label2Label is conceptually simple and empirically powerful. Without incorporating task-specific prior knowledge and highly specialized network designs, our approach achieves state-of-the-art results on three different multi-attribute learning tasks, compared to highly customized domain-specific methods. Code is available at https://github.com/Li-Wanhua/Label2Label.

Ziyi Wang, Xumin Yu, Yongming Rao et al. · tsinghua

With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNetPart segmentation tasks. Code is available at https://github.com/wangzy22/TAP.

Xiuwei Xu, Yifan Wang, Yu Zheng et al. · tsinghua

In this paper, we propose a weakly-supervised approach for 3D object detection, which makes it possible to train a strong 3D detector with position-level annotations (i.e. annotations of object centers). In order to remedy the information loss from box annotations to centers, our method, namely Back to Reality (BR), makes use of synthetic 3D shapes to convert the weak labels into fully-annotated virtual scenes as stronger supervision, and in turn utilizes the perfect virtual labels to complement and refine the real labels. Specifically, we first assemble 3D shapes into physically reasonable virtual scenes according to the coarse scene layout extracted from position-level annotations. Then we go back to reality by applying a virtual-to-real domain adaptation method, which refine the weak labels and additionally supervise the training of detector with the virtual scenes. Furthermore, we propose a more challenging benckmark for indoor 3D object detection with more diversity in object sizes to better show the potential of BR. With less than 5% of the labeling labor, we achieve comparable detection performance with some popular fully-supervised approaches on the widely used ScanNet dataset. Code is available at: https://github.com/wyf-ACCEPT/BackToReality

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.

Junlong Li, Bingyao Yu, Yongming Rao et al. · tencent-ai, tsinghua

In recent years, significant progress has been made in video instance segmentation (VIS), with many offline and online methods achieving state-of-the-art performance. While offline methods have the advantage of producing temporally consistent predictions, they are not suitable for real-time scenarios. Conversely, online methods are more practical, but maintaining temporal consistency remains a challenging task. In this paper, we propose a novel online method for video instance segmentation, called TCOVIS, which fully exploits the temporal information in a video clip. The core of our method consists of a global instance assignment strategy and a spatio-temporal enhancement module, which improve the temporal consistency of the features from two aspects. Specifically, we perform global optimal matching between the predictions and ground truth across the whole video clip, and supervise the model with the global optimal objective. We also capture the spatial feature and aggregate it with the semantic feature between frames, thus realizing the spatio-temporal enhancement. We evaluate our method on four widely adopted VIS benchmarks, namely YouTube-VIS 2019/2021/2022 and OVIS, and achieve state-of-the-art performance on all benchmarks without bells-and-whistles. For instance, on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. Code is available at https://github.com/jun-long-li/TCOVIS.

Hui Li, Tianyang Xu, Xiao-Jun Wu et al.

Deep learning based fusion methods have been achieving promising performance in image fusion tasks. This is attributed to the network architecture that plays a very important role in the fusion process. However, in general, it is hard to specify a good fusion architecture, and consequently, the design of fusion networks is still a black art, rather than science. To address this problem, we formulate the fusion task mathematically, and establish a connection between its optimal solution and the network architecture that can implement it. This approach leads to a novel method proposed in the paper of constructing a lightweight fusion network. It avoids the time-consuming empirical network design by a trial-and-test strategy. In particular we adopt a learnable representation approach to the fusion task, in which the construction of the fusion network architecture is guided by the optimisation algorithm producing the learnable model. The low-rank representation (LRR) objective is the foundation of our learnable model. The matrix multiplications, which are at the heart of the solution are transformed into convolutional operations, and the iterative process of optimisation is replaced by a special feed-forward network. Based on this novel network architecture, an end-to-end lightweight fusion network is constructed to fuse infrared and visible light images. Its successful training is facilitated by a detail-to-semantic information loss function proposed to preserve the image details and to enhance the salient features of the source images. Our experiments show that the proposed fusion network exhibits better fusion performance than the state-of-the-art fusion methods on public datasets. Interestingly, our network requires a fewer training parameters than other existing methods. The codes are available at https://github.com/hli1221/imagefusion-LRRNet

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.

Borui Zhang, Wenzhao Zheng, Jie Zhou et al. · tsinghua

This paper proposes an attributable visual similarity learning (AVSL) framework for a more accurate and explainable similarity measure between images. Most existing similarity learning methods exacerbate the unexplainability by mapping each sample to a single point in the embedding space with a distance metric (e.g., Mahalanobis distance, Euclidean distance). Motivated by the human semantic similarity cognition, we propose a generalized similarity learning paradigm to represent the similarity between two images with a graph and then infer the overall similarity accordingly. Furthermore, we establish a bottom-up similarity construction and top-down similarity inference framework to infer the similarity based on semantic hierarchy consistency. We first identify unreliable higher-level similarity nodes and then correct them using the most coherent adjacent lower-level similarity nodes, which simultaneously preserve traces for similarity attribution. Extensive experiments on the CUB-200-2011, Cars196, and Stanford Online Products datasets demonstrate significant improvements over existing deep similarity learning methods and verify the interpretability of our framework. Code is available at https://github.com/zbr17/AVSL.

Ziyi Wang, Yongming Rao, Xumin Yu et al. · tsinghua

Conventional point cloud semantic segmentation methods usually employ an encoder-decoder architecture, where mid-level features are locally aggregated to extract geometric information. However, the over-reliance on these class-agnostic local geometric representations may raise confusion between local parts from different categories that are similar in appearance or spatially adjacent. To address this issue, we argue that mid-level features can be further enhanced with semantic information, and propose semantic-affine transformation that transforms features of mid-level points belonging to different categories with class-specific affine parameters. Based on this technique, we propose SemAffiNet for point cloud semantic segmentation, which utilizes the attention mechanism in the Transformer module to implicitly and explicitly capture global structural knowledge within local parts for overall comprehension of each category. We conduct extensive experiments on the ScanNetV2 and NYUv2 datasets, and evaluate semantic-affine transformation on various 3D point cloud and 2D image segmentation baselines, where both qualitative and quantitative results demonstrate the superiority and generalization ability of our proposed approach. Code is available at https://github.com/wangzy22/SemAffiNet.

Zuyan Liu, Yuhao Dong, Ziwei Liu et al. · tsinghua

Visual data comes in various forms, ranging from small icons of just a few pixels to long videos spanning hours. Existing multi-modal LLMs usually standardize these diverse visual inputs to a fixed resolution for visual encoders and yield similar numbers of tokens for LLMs. This approach is non-optimal for multimodal understanding and inefficient for processing inputs with long and short visual contents. To solve the problem, we propose Oryx, a unified multimodal architecture for the spatial-temporal understanding of images, videos, and multi-view 3D scenes. Oryx offers an on-demand solution to seamlessly and efficiently process visual inputs with arbitrary spatial sizes and temporal lengths through two core innovations: 1) a pre-trained OryxViT model that can encode images at any resolution into LLM-friendly visual representations; 2) a dynamic compressor module that supports 1x to 16x compression on visual tokens by request. These design features enable Oryx to accommodate extremely long visual contexts, such as videos, with lower resolution and high compression while maintaining high recognition precision for tasks like document understanding with native resolution and no compression. Beyond the architectural improvements, enhanced data curation and specialized training on long-context retrieval and spatial-aware data help Oryx achieve strong capabilities in image, video, and 3D multimodal understanding simultaneously. Our work is open-sourced at https://github.com/Oryx-mllm/Oryx.

Chengkun Wang, Wenzhao Zheng, Xian Sun et al. · tsinghua

This paper proposes a probabilistic deep metric learning (PDML) framework for hyperspectral image classification, which aims to predict the category of each pixel for an image captured by hyperspectral sensors. The core problem for hyperspectral image classification is the spectral variability between intraclass materials and the spectral similarity between interclass materials, motivating the further incorporation of spatial information to differentiate a pixel based on its surrounding patch. However, different pixels and even the same pixel in one patch might not encode the same material due to the low spatial resolution of most hyperspectral sensors, leading to an inconsistent judgment of a specific pixel. To address this issue, we propose a probabilistic deep metric learning framework to model the categorical uncertainty of the spectral distribution of an observed pixel. We propose to learn a global probabilistic distribution for each pixel in the patch and a probabilistic metric to model the distance between distributions. We treat each pixel in a patch as a training sample, enabling us to exploit more information from the patch compared with conventional methods. Our framework can be readily applied to existing hyperspectral image classification methods with various network architectures and loss functions. Extensive experiments on four widely used datasets including IN, UP, KSC, and Houston 2013 datasets demonstrate that our framework improves the performance of existing methods and further achieves the state of the art. Code is available at: https://github.com/wzzheng/PDML.

An Tao, Yueqi Duan, Yingqi Wang et al. · tsinghua

In this paper, we investigate the dynamics-aware adversarial attack problem of adaptive neural networks. Most existing adversarial attack algorithms are designed under a basic assumption -- the network architecture is fixed throughout the attack process. However, this assumption does not hold for many recently proposed adaptive neural networks, which adaptively deactivate unnecessary execution units based on inputs to improve computational efficiency. It results in a serious issue of lagged gradient, making the learned attack at the current step ineffective due to the architecture change afterward. To address this issue, we propose a Leaded Gradient Method (LGM) and show the significant effects of the lagged gradient. More specifically, we reformulate the gradients to be aware of the potential dynamic changes of network architectures, so that the learned attack better "leads" the next step than the dynamics-unaware methods when network architecture changes dynamically. Extensive experiments on representative types of adaptive neural networks for both 2D images and 3D point clouds show that our LGM achieves impressive adversarial attack performance compared with the dynamic-unaware attack methods. Code is available at https://github.com/antao97/LGM.

Borui Zhang, Wenzhao Zheng, Jie Zhou et al. · tsinghua

Deep learning has revolutionized human society, yet the black-box nature of deep neural networks hinders further application to reliability-demanded industries. In the attempt to unpack them, many works observe or impact internal variables to improve the comprehensibility and invertibility of the black-box models. However, existing methods rely on intuitive assumptions and lack mathematical guarantees. To bridge this gap, we introduce Bort, an optimizer for improving model explainability with boundedness and orthogonality constraints on model parameters, derived from the sufficient conditions of model comprehensibility and invertibility. We perform reconstruction and backtracking on the model representations optimized by Bort and observe a clear improvement in model explainability. Based on Bort, we are able to synthesize explainable adversarial samples without additional parameters and training. Surprisingly, we find Bort constantly improves the classification accuracy of various architectures including ResNet and DeiT on MNIST, CIFAR-10, and ImageNet. Code: https://github.com/zbr17/Bort.

Wenzhao Zheng, Weiliang Chen, Yuanhui Huang et al. · tsinghua

Understanding how the 3D scene evolves is vital for making decisions in autonomous driving. Most existing methods achieve this by predicting the movements of object boxes, which cannot capture more fine-grained scene information. In this paper, we explore a new framework of learning a world model, OccWorld, in the 3D Occupancy space to simultaneously predict the movement of the ego car and the evolution of the surrounding scenes. We propose to learn a world model based on 3D occupancy rather than 3D bounding boxes and segmentation maps for three reasons: 1) expressiveness. 3D occupancy can describe the more fine-grained 3D structure of the scene; 2) efficiency. 3D occupancy is more economical to obtain (e.g., from sparse LiDAR points). 3) versatility. 3D occupancy can adapt to both vision and LiDAR. To facilitate the modeling of the world evolution, we learn a reconstruction-based scene tokenizer on the 3D occupancy to obtain discrete scene tokens to describe the surrounding scenes. We then adopt a GPT-like spatial-temporal generative transformer to generate subsequent scene and ego tokens to decode the future occupancy and ego trajectory. Extensive experiments on the widely used nuScenes benchmark demonstrate the ability of OccWorld to effectively model the evolution of the driving scenes. OccWorld also produces competitive planning results without using instance and map supervision. Code: https://github.com/wzzheng/OccWorld.

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.

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

The pretrain-finetune paradigm in modern computer vision facilitates the success of self-supervised learning, which tends to achieve better transferability than supervised learning. However, with the availability of massive labeled data, a natural question emerges: how to train a better model with both self and full supervision signals? In this paper, we propose Omni-suPErvised Representation leArning with hierarchical supervisions (OPERA) as a solution. We provide a unified perspective of supervisions from labeled and unlabeled data and propose a unified framework of fully supervised and self-supervised learning. We extract a set of hierarchical proxy representations for each image and impose self and full supervisions on the corresponding proxy representations. Extensive experiments on both convolutional neural networks and vision transformers demonstrate the superiority of OPERA in image classification, segmentation, and object detection. Code is available at: https://github.com/wangck20/OPERA.

Muheng Li, Lei Chen, Yueqi Duan et al.

Action recognition models have shown a promising capability to classify human actions in short video clips. In a real scenario, multiple correlated human actions commonly occur in particular orders, forming semantically meaningful human activities. Conventional action recognition approaches focus on analyzing single actions. However, they fail to fully reason about the contextual relations between adjacent actions, which provide potential temporal logic for understanding long videos. In this paper, we propose a prompt-based framework, Bridge-Prompt (Br-Prompt), to model the semantics across adjacent actions, so that it simultaneously exploits both out-of-context and contextual information from a series of ordinal actions in instructional videos. More specifically, we reformulate the individual action labels as integrated text prompts for supervision, which bridge the gap between individual action semantics. The generated text prompts are paired with corresponding video clips, and together co-train the text encoder and the video encoder via a contrastive approach. The learned vision encoder has a stronger capability for ordinal-action-related downstream tasks, e.g. action segmentation and human activity recognition. We evaluate the performances of our approach on several video datasets: Georgia Tech Egocentric Activities (GTEA), 50Salads, and the Breakfast dataset. Br-Prompt achieves state-of-the-art on multiple benchmarks. Code is available at https://github.com/ttlmh/Bridge-Prompt

Zuyan Liu, Benlin Liu, Jiahui Wang et al. · tsinghua

In the field of instruction-following large vision-language models (LVLMs), the efficient deployment of these models faces challenges, notably due to the high memory demands of their key-value (KV) caches. Conventional cache management strategies for LLMs focus on cache eviction, which often fails to address the specific needs of multimodal instruction-following models. Recognizing this gap, in this paper, we introduce Elastic Cache, a novel approach that benefits from applying distinct acceleration methods for instruction encoding and output generation stages. We investigate the metrics of importance in different stages and propose an importance-driven cache merging strategy to prune redundancy caches. Instead of discarding less important caches, our strategy identifies important key/value vectors as anchor points. Surrounding less important caches are then merged with these anchors, enhancing the preservation of contextual information in the KV caches while yielding an arbitrary acceleration ratio. For instruction encoding, we utilize the frequency to evaluate the importance of caches. Regarding output generation, we prioritize tokens based on their distance with an offset, by which both the initial and most recent tokens are retained. Results on a range of LVLMs demonstrate that Elastic Cache not only boosts efficiency but also notably outperforms existing pruning methods in language generation across various tasks. Code is available at https://github.com/liuzuyan/ElasticCache

Shuai Shen, Wenliang Zhao, Zibin Meng et al. · tsinghua

Talking head synthesis is a promising approach for the video production industry. Recently, a lot of effort has been devoted in this research area to improve the generation quality or enhance the model generalization. However, there are few works able to address both issues simultaneously, which is essential for practical applications. To this end, in this paper, we turn attention to the emerging powerful Latent Diffusion Models, and model the Talking head generation as an audio-driven temporally coherent denoising process (DiffTalk). More specifically, instead of employing audio signals as the single driving factor, we investigate the control mechanism of the talking face, and incorporate reference face images and landmarks as conditions for personality-aware generalized synthesis. In this way, the proposed DiffTalk is capable of producing high-quality talking head videos in synchronization with the source audio, and more importantly, it can be naturally generalized across different identities without any further fine-tuning. Additionally, our DiffTalk can be gracefully tailored for higher-resolution synthesis with negligible extra computational cost. Extensive experiments show that the proposed DiffTalk efficiently synthesizes high-fidelity audio-driven talking head videos for generalized novel identities. For more video results, please refer to \url{https://sstzal.github.io/DiffTalk/}.

Yansong Tang, Xingyu Liu, Xumin Yu et al.

Rapid progress and superior performance have been achieved for skeleton-based action recognition recently. In this article, we investigate this problem under a cross-dataset setting, which is a new, pragmatic, and challenging task in real-world scenarios. Following the unsupervised domain adaptation (UDA) paradigm, the action labels are only available on a source dataset, but unavailable on a target dataset in the training stage. Different from the conventional adversarial learning-based approaches for UDA, we utilize a self-supervision scheme to reduce the domain shift between two skeleton-based action datasets. Our inspiration is drawn from Cubism, an art genre from the early 20th century, which breaks and reassembles the objects to convey a greater context. By segmenting and permuting temporal segments or human body parts, we design two self-supervised learning classification tasks to explore the temporal and spatial dependency of a skeleton-based action and improve the generalization ability of the model. We conduct experiments on six datasets for skeleton-based action recognition, including three large-scale datasets (NTU RGB+D, PKU-MMD, and Kinetics) where new cross-dataset settings and benchmarks are established. Extensive results demonstrate that our method outperforms state-of-the-art approaches. The source codes of our model and all the compared methods are available at https://github.com/shanice-l/st-cubism.

Wenzhao Zheng, Chengkun Wang, Jie Zhou et al. · tsinghua

This paper proposes an introspective deep metric learning (IDML) framework for uncertainty-aware comparisons of images. Conventional deep metric learning methods produce confident semantic distances between images regardless of the uncertainty level. However, we argue that a good similarity model should consider the semantic discrepancies with caution 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 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 is available at: https://github.com/wzzheng/IDML.

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.

Yi Wei, Zibu Wei, Yongming Rao et al. · tsinghua

In this paper, we propose the LiDAR Distillation to bridge the domain gap induced by different LiDAR beams for 3D object detection. In many real-world applications, the LiDAR points used by mass-produced robots and vehicles usually have fewer beams than that in large-scale public datasets. Moreover, as the LiDARs are upgraded to other product models with different beam amount, it becomes challenging to utilize the labeled data captured by previous versions' high-resolution sensors. Despite the recent progress on domain adaptive 3D detection, most methods struggle to eliminate the beam-induced domain gap. We find that it is essential to align the point cloud density of the source domain with that of the target domain during the training process. Inspired by this discovery, we propose a progressive framework to mitigate the beam-induced domain shift. In each iteration, we first generate low-beam pseudo LiDAR by downsampling the high-beam point clouds. Then the teacher-student framework is employed to distill rich information from the data with more beams. Extensive experiments on Waymo, nuScenes and KITTI datasets with three different LiDAR-based detectors demonstrate the effectiveness of our LiDAR Distillation. Notably, our approach does not increase any additional computation cost for inference.

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.

Wenliang Zhao, Haolin Wang, Jie Zhou et al.

Diffusion probabilistic models (DPMs) have shown remarkable performance in visual synthesis but are computationally expensive due to the need for multiple evaluations during the sampling. Recent predictor-corrector diffusion samplers have significantly reduced the required number of function evaluations (NFE), but inherently suffer from a misalignment issue caused by the extra corrector step, especially with a large classifier-free guidance scale (CFG). In this paper, we introduce a new fast DPM sampler called DC-Solver, which leverages dynamic compensation (DC) to mitigate the misalignment of the predictor-corrector samplers. The dynamic compensation is controlled by compensation ratios that are adaptive to the sampling steps and can be optimized on only 10 datapoints by pushing the sampling trajectory toward a ground truth trajectory. We further propose a cascade polynomial regression (CPR) which can instantly predict the compensation ratios on unseen sampling configurations. Additionally, we find that the proposed dynamic compensation can also serve as a plug-and-play module to boost the performance of predictor-only samplers. Extensive experiments on both unconditional sampling and conditional sampling demonstrate that our DC-Solver can consistently improve the sampling quality over previous methods on different DPMs with a wide range of resolutions up to 1024$\times$1024. Notably, we achieve 10.38 FID (NFE=5) on unconditional FFHQ and 0.394 MSE (NFE=5, CFG=7.5) on Stable-Diffusion-2.1. Code is available at https://github.com/wl-zhao/DC-Solver

Quan Zheng, Ziwei Wang, Jie Zhou et al. · tsinghua

Explaining deep convolutional neural networks has been recently drawing increasing attention since it helps to understand the networks' internal operations and why they make certain decisions. Saliency maps, which emphasize salient regions largely connected to the network's decision-making, are one of the most common ways for visualizing and analyzing deep networks in the computer vision community. However, saliency maps generated by existing methods cannot represent authentic information in images due to the unproven proposals about the weights of activation maps which lack solid theoretical foundation and fail to consider the relations between each pixel. In this paper, we develop a novel post-hoc visual explanation method called Shap-CAM based on class activation mapping. Unlike previous gradient-based approaches, Shap-CAM gets rid of the dependence on gradients by obtaining the importance of each pixel through Shapley value. We demonstrate that Shap-CAM achieves better visual performance and fairness for interpreting the decision making process. Our approach outperforms previous methods on both recognition and localization tasks.

Yansong Tang, Jinpeng Liu, Aoyang Liu et al. · tsinghua

With the continuously thriving popularity around the world, fitness activity analytic has become an emerging research topic in computer vision. While a variety of new tasks and algorithms have been proposed recently, there are growing hunger for data resources involved in high-quality data, fine-grained labels, and diverse environments. In this paper, we present FLAG3D, a large-scale 3D fitness activity dataset with language instruction containing 180K sequences of 60 categories. FLAG3D features the following three aspects: 1) accurate and dense 3D human pose captured from advanced MoCap system to handle the complex activity and large movement, 2) detailed and professional language instruction to describe how to perform a specific activity, 3) versatile video resources from a high-tech MoCap system, rendering software, and cost-effective smartphones in natural environments. Extensive experiments and in-depth analysis show that FLAG3D contributes great research value for various challenges, such as cross-domain human action recognition, dynamic human mesh recovery, and language-guided human action generation. Our dataset and source code are publicly available at https://andytang15.github.io/FLAG3D.

Wenliang Zhao, Minglei Shi, Xumin Yu et al.

Building on the success of diffusion models in visual generation, flow-based models reemerge as another prominent family of generative models that have achieved competitive or better performance in terms of both visual quality and inference speed. By learning the velocity field through flow-matching, flow-based models tend to produce a straighter sampling trajectory, which is advantageous during the sampling process. However, unlike diffusion models for which fast samplers are well-developed, efficient sampling of flow-based generative models has been rarely explored. In this paper, we propose a framework called FlowTurbo to accelerate the sampling of flow-based models while still enhancing the sampling quality. Our primary observation is that the velocity predictor's outputs in the flow-based models will become stable during the sampling, enabling the estimation of velocity via a lightweight velocity refiner. Additionally, we introduce several techniques including a pseudo corrector and sample-aware compilation to further reduce inference time. Since FlowTurbo does not change the multi-step sampling paradigm, it can be effectively applied for various tasks such as image editing, inpainting, etc. By integrating FlowTurbo into different flow-based models, we obtain an acceleration ratio of 53.1%$\sim$58.3% on class-conditional generation and 29.8%$\sim$38.5% on text-to-image generation. Notably, FlowTurbo reaches an FID of 2.12 on ImageNet with 100 (ms / img) and FID of 3.93 with 38 (ms / img), achieving the real-time image generation and establishing the new state-of-the-art. Code is available at https://github.com/shiml20/FlowTurbo.

Yu Zheng, Yueqi Duan, Jiwen Lu et al. · tsinghua

A bathtub in a library, a sink in an office, a bed in a laundry room -- the counter-intuition suggests that scene provides important prior knowledge for 3D object detection, which instructs to eliminate the ambiguous detection of similar objects. In this paper, we propose HyperDet3D to explore scene-conditioned prior knowledge for 3D object detection. Existing methods strive for better representation of local elements and their relations without scene-conditioned knowledge, which may cause ambiguity merely based on the understanding of individual points and object candidates. Instead, HyperDet3D simultaneously learns scene-agnostic embeddings and scene-specific knowledge through scene-conditioned hypernetworks. More specifically, our HyperDet3D not only explores the sharable abstracts from various 3D scenes, but also adapts the detector to the given scene at test time. We propose a discriminative Multi-head Scene-specific Attention (MSA) module to dynamically control the layer parameters of the detector conditioned on the fusion of scene-conditioned knowledge. Our HyperDet3D achieves state-of-the-art results on the 3D object detection benchmark of the ScanNet and SUN RGB-D datasets. Moreover, through cross-dataset evaluation, we show the acquired scene-conditioned prior knowledge still takes effect when facing 3D scenes with domain gap.

Sichao Huang, Ziwei Wang, Jie Zhou et al. · tsinghua

Object packing by autonomous robots is an im-portant challenge in warehouses and logistics industry. Most conventional data-driven packing planning approaches focus on regular cuboid packing, which are usually heuristic and limit the practical use in realistic applications with everyday objects. In this paper, we propose a deep hierarchical reinforcement learning approach to simultaneously plan packing sequence and placement for irregular object packing. Specifically, the top manager network infers packing sequence from six principal view heightmaps of all objects, and then the bottom worker network receives heightmaps of the next object to predict the placement position and orientation. The two networks are trained hierarchically in a self-supervised Q-Learning framework, where the rewards are provided by the packing results based on the top height , object volume and placement stability in the box. The framework repeats sequence and placement planning iteratively until all objects have been packed into the box or no space is remained for unpacked items. We compare our approach with existing robotic packing methods for irregular objects in a physics simulator. Experiments show that our approach can pack more objects with less time cost than the state-of-the-art packing methods of irregular objects. We also implement our packing plan with a robotic manipulator to show the generalization ability in the real world.

Sha Yuan, Hanyu Zhao, Shuai Zhao et al. · bytedance, pku

With the rapid development of deep learning, training Big Models (BMs) for multiple downstream tasks becomes a popular paradigm. Researchers have achieved various outcomes in the construction of BMs and the BM application in many fields. At present, there is a lack of research work that sorts out the overall progress of BMs and guides the follow-up research. In this paper, we cover not only the BM technologies themselves but also the prerequisites for BM training and applications with BMs, dividing the BM review into four parts: Resource, Models, Key Technologies and Application. We introduce 16 specific BM-related topics in those four parts, they are Data, Knowledge, Computing System, Parallel Training System, Language Model, Vision Model, Multi-modal Model, Theory&Interpretability, Commonsense Reasoning, Reliability&Security, Governance, Evaluation, Machine Translation, Text Generation, Dialogue and Protein Research. In each topic, we summarize clearly the current studies and propose some future research directions. At the end of this paper, we conclude the further development of BMs in a more general view.

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.