Yuenan Hou

Xin Li, Tao Ma, Yuenan Hou et al. · stanford

LiDAR-camera fusion methods have shown impressive performance in 3D object detection. Recent advanced multi-modal methods mainly perform global fusion, where image features and point cloud features are fused across the whole scene. Such practice lacks fine-grained region-level information, yielding suboptimal fusion performance. In this paper, we present the novel Local-to-Global fusion network (LoGoNet), which performs LiDAR-camera fusion at both local and global levels. Concretely, the Global Fusion (GoF) of LoGoNet is built upon previous literature, while we exclusively use point centroids to more precisely represent the position of voxel features, thus achieving better cross-modal alignment. As to the Local Fusion (LoF), we first divide each proposal into uniform grids and then project these grid centers to the images. The image features around the projected grid points are sampled to be fused with position-decorated point cloud features, maximally utilizing the rich contextual information around the proposals. The Feature Dynamic Aggregation (FDA) module is further proposed to achieve information interaction between these locally and globally fused features, thus producing more informative multi-modal features. Extensive experiments on both Waymo Open Dataset (WOD) and KITTI datasets show that LoGoNet outperforms all state-of-the-art 3D detection methods. Notably, LoGoNet ranks 1st on Waymo 3D object detection leaderboard and obtains 81.02 mAPH (L2) detection performance. It is noteworthy that, for the first time, the detection performance on three classes surpasses 80 APH (L2) simultaneously. Code will be available at \url{https://github.com/sankin97/LoGoNet}.

Yuenan Hou, Xinge Zhu, Yuexin Ma et al.

This article addresses the problem of distilling knowledge from a large teacher model to a slim student network for LiDAR semantic segmentation. Directly employing previous distillation approaches yields inferior results due to the intrinsic challenges of point cloud, i.e., sparsity, randomness and varying density. To tackle the aforementioned problems, we propose the Point-to-Voxel Knowledge Distillation (PVD), which transfers the hidden knowledge from both point level and voxel level. Specifically, we first leverage both the pointwise and voxelwise output distillation to complement the sparse supervision signals. Then, to better exploit the structural information, we divide the whole point cloud into several supervoxels and design a difficulty-aware sampling strategy to more frequently sample supervoxels containing less-frequent classes and faraway objects. On these supervoxels, we propose inter-point and inter-voxel affinity distillation, where the similarity information between points and voxels can help the student model better capture the structural information of the surrounding environment. We conduct extensive experiments on two popular LiDAR segmentation benchmarks, i.e., nuScenes and SemanticKITTI. On both benchmarks, our PVD consistently outperforms previous distillation approaches by a large margin on three representative backbones, i.e., Cylinder3D, SPVNAS and MinkowskiNet. Notably, on the challenging nuScenes and SemanticKITTI datasets, our method can achieve roughly 75% MACs reduction and 2x speedup on the competitive Cylinder3D model and rank 1st on the SemanticKITTI leaderboard among all published algorithms. Our code is available at https://github.com/cardwing/Codes-for-PVKD.

Runnan Chen, Youquan Liu, Lingdong Kong et al.

Contrastive Language-Image Pre-training (CLIP) achieves promising results in 2D zero-shot and few-shot learning. Despite the impressive performance in 2D, applying CLIP to help the learning in 3D scene understanding has yet to be explored. In this paper, we make the first attempt to investigate how CLIP knowledge benefits 3D scene understanding. We propose CLIP2Scene, a simple yet effective framework that transfers CLIP knowledge from 2D image-text pre-trained models to a 3D point cloud network. We show that the pre-trained 3D network yields impressive performance on various downstream tasks, i.e., annotation-free and fine-tuning with labelled data for semantic segmentation. Specifically, built upon CLIP, we design a Semantic-driven Cross-modal Contrastive Learning framework that pre-trains a 3D network via semantic and spatial-temporal consistency regularization. For the former, we first leverage CLIP's text semantics to select the positive and negative point samples and then employ the contrastive loss to train the 3D network. In terms of the latter, we force the consistency between the temporally coherent point cloud features and their corresponding image features. We conduct experiments on SemanticKITTI, nuScenes, and ScanNet. For the first time, our pre-trained network achieves annotation-free 3D semantic segmentation with 20.8% and 25.08% mIoU on nuScenes and ScanNet, respectively. When fine-tuned with 1% or 100% labelled data, our method significantly outperforms other self-supervised methods, with improvements of 8% and 1% mIoU, respectively. Furthermore, we demonstrate the generalizability for handling cross-domain datasets. Code is publicly available https://github.com/runnanchen/CLIP2Scene.

Xiaoshui Huang, Zhou Huang, Sheng Li et al.

The pretrain-finetune paradigm has achieved great success in NLP and 2D image fields because of the high-quality representation ability and transferability of their pretrained models. However, pretraining such a strong model is difficult in the 3D point cloud field due to the limited amount of point cloud sequences. This paper introduces \textbf{E}fficient \textbf{P}oint \textbf{C}loud \textbf{L}earning (EPCL), an effective and efficient point cloud learner for directly training high-quality point cloud models with a frozen CLIP transformer. Our EPCL connects the 2D and 3D modalities by semantically aligning the image features and point cloud features without paired 2D-3D data. Specifically, the input point cloud is divided into a series of local patches, which are converted to token embeddings by the designed point cloud tokenizer. These token embeddings are concatenated with a task token and fed into the frozen CLIP transformer to learn point cloud representation. The intuition is that the proposed point cloud tokenizer projects the input point cloud into a unified token space that is similar to the 2D images. Comprehensive experiments on 3D detection, semantic segmentation, classification and few-shot learning demonstrate that the CLIP transformer can serve as an efficient point cloud encoder and our method achieves promising performance on both indoor and outdoor benchmarks. In particular, performance gains brought by our EPCL are $\textbf{19.7}$ AP$_{50}$ on ScanNet V2 detection, $\textbf{4.4}$ mIoU on S3DIS segmentation and $\textbf{1.2}$ mIoU on SemanticKITTI segmentation compared to contemporary pretrained models. Code is available at \url{https://github.com/XiaoshuiHuang/EPCL}.

Youquan Liu, Runnan Chen, Xin Li et al.

Point-, voxel-, and range-views are three representative forms of point clouds. All of them have accurate 3D measurements but lack color and texture information. RGB images are a natural complement to these point cloud views and fully utilizing the comprehensive information of them benefits more robust perceptions. In this paper, we present a unified multi-modal LiDAR segmentation network, termed UniSeg, which leverages the information of RGB images and three views of the point cloud, and accomplishes semantic segmentation and panoptic segmentation simultaneously. Specifically, we first design the Learnable cross-Modal Association (LMA) module to automatically fuse voxel-view and range-view features with image features, which fully utilize the rich semantic information of images and are robust to calibration errors. Then, the enhanced voxel-view and range-view features are transformed to the point space,where three views of point cloud features are further fused adaptively by the Learnable cross-View Association module (LVA). Notably, UniSeg achieves promising results in three public benchmarks, i.e., SemanticKITTI, nuScenes, and Waymo Open Dataset (WOD); it ranks 1st on two challenges of two benchmarks, including the LiDAR semantic segmentation challenge of nuScenes and panoptic segmentation challenges of SemanticKITTI. Besides, we construct the OpenPCSeg codebase, which is the largest and most comprehensive outdoor LiDAR segmentation codebase. It contains most of the popular outdoor LiDAR segmentation algorithms and provides reproducible implementations. The OpenPCSeg codebase will be made publicly available at https://github.com/PJLab-ADG/PCSeg.

Zhenxiang Lin, Xidong Peng, Peishan Cong et al.

We introduce the task of 3D visual grounding in large-scale dynamic scenes based on natural linguistic descriptions and online captured multi-modal visual data, including 2D images and 3D LiDAR point clouds. We present a novel method, dubbed WildRefer, for this task by fully utilizing the rich appearance information in images, the position and geometric clues in point cloud as well as the semantic knowledge of language descriptions. Besides, we propose two novel datasets, i.e., STRefer and LifeRefer, which focus on large-scale human-centric daily-life scenarios accompanied with abundant 3D object and natural language annotations. Our datasets are significant for the research of 3D visual grounding in the wild and has huge potential to boost the development of autonomous driving and service robots. Extensive experiments and ablation studies demonstrate that our method achieves state-of-the-art performance on the proposed benchmarks. The code is provided in https://github.com/4DVLab/WildRefer.

Yuexin Ma, Tai Wang, Xuyang Bai et al.

In recent years, vision-centric Bird's Eye View (BEV) perception has garnered significant interest from both industry and academia due to its inherent advantages, such as providing an intuitive representation of the world and being conducive to data fusion. The rapid advancements in deep learning have led to the proposal of numerous methods for addressing vision-centric BEV perception challenges. However, there has been no recent survey encompassing this novel and burgeoning research field. To catalyze future research, this paper presents a comprehensive survey of the latest developments in vision-centric BEV perception and its extensions. It compiles and organizes up-to-date knowledge, offering a systematic review and summary of prevalent algorithms. Additionally, the paper provides in-depth analyses and comparative results on various BEV perception tasks, facilitating the evaluation of future works and sparking new research directions. Furthermore, the paper discusses and shares valuable empirical implementation details to aid in the advancement of related algorithms.

Xiaopei Wu, Yuenan Hou, Xiaoshui Huang et al.

Training deep models for LiDAR semantic segmentation is challenging due to the inherent sparsity of point clouds. Utilizing temporal data is a natural remedy against the sparsity problem as it makes the input signal denser. However, previous multi-frame fusion algorithms fall short in utilizing sufficient temporal information due to the memory constraint, and they also ignore the informative temporal images. To fully exploit rich information hidden in long-term temporal point clouds and images, we present the Temporal Aggregation Network, termed TASeg. Specifically, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm, which leverages historical priors to assign different aggregation steps for different classes. It can largely reduce memory and time overhead while achieving higher accuracy. Besides, TLAD trains a teacher injected with gt priors to distill the model, further boosting the performance. To make full use of temporal images, we design a Temporal Image Aggregation and Fusion (TIAF) module, which can greatly expand the camera FOV and enhance the present features. Temporal LiDAR points in the camera FOV are used as mediums to transform temporal image features to the present coordinate for temporal multi-modal fusion. Moreover, we develop a Static-Moving Switch Augmentation (SMSA) algorithm, which utilizes sufficient temporal information to enable objects to switch their motion states freely, thus greatly increasing static and moving training samples. Our TASeg ranks 1st on three challenging tracks, i.e., SemanticKITTI single-scan track, multi-scan track and nuScenes LiDAR segmentation track, strongly demonstrating the superiority of our method. Codes are available at https://github.com/LittlePey/TASeg.

Peishan Cong, Xinge Zhu, Feng Qiao et al.

Accurately detecting and tracking pedestrians in 3D space is challenging due to large variations in rotations, poses and scales. The situation becomes even worse for dense crowds with severe occlusions. However, existing benchmarks either only provide 2D annotations, or have limited 3D annotations with low-density pedestrian distribution, making it difficult to build a reliable pedestrian perception system especially in crowded scenes. To better evaluate pedestrian perception algorithms in crowded scenarios, we introduce a large-scale multimodal dataset,STCrowd. Specifically, in STCrowd, there are a total of 219 K pedestrian instances and 20 persons per frame on average, with various levels of occlusion. We provide synchronized LiDAR point clouds and camera images as well as their corresponding 3D labels and joint IDs. STCrowd can be used for various tasks, including LiDAR-only, image-only, and sensor-fusion based pedestrian detection and tracking. We provide baselines for most of the tasks. In addition, considering the property of sparse global distribution and density-varying local distribution of pedestrians, we further propose a novel method, Density-aware Hierarchical heatmap Aggregation (DHA), to enhance pedestrian perception in crowded scenes. Extensive experiments show that our new method achieves state-of-the-art performance for pedestrian detection on various datasets.

Weijie Li, Yang Wei, Tianpeng Liu et al.

The growing Synthetic Aperture Radar (SAR) data has the potential to build a foundation model through Self-Supervised Learning (SSL) methods, which can achieve various SAR Automatic Target Recognition (ATR) tasks with pre-training in large-scale unlabeled data and fine-tuning in small labeled samples. SSL aims to construct supervision signals directly from the data, which minimizes the need for expensive expert annotation and maximizes the use of the expanding data pool for a foundational model. This study investigates an effective SSL method for SAR ATR, which can pave the way for a foundation model in SAR ATR. The primary obstacles faced in SSL for SAR ATR are the small targets in remote sensing and speckle noise in SAR images, corresponding to the SSL approach and signals. To overcome these challenges, we present a novel Joint-Embedding Predictive Architecture for SAR ATR (SAR-JEPA), which leverages local masked patches to predict the multi-scale SAR gradient representations of unseen context. The key aspect of SAR-JEPA is integrating SAR domain features to ensure high-quality self-supervised signals as target features. Besides, we employ local masks and multi-scale features to accommodate the various small targets in remote sensing. By fine-tuning and evaluating our framework on three target recognition datasets (vehicle, ship, and aircraft) with four other datasets as pre-training, we demonstrate its outperformance over other SSL methods and its effectiveness with increasing SAR data. This study showcases the potential of SSL for SAR target recognition across diverse targets, scenes, and sensors.Our codes and weights are available in \url{https://github.com/waterdisappear/SAR-JEPA.

Lingdong Kong, Youquan Liu, Runnan Chen et al.

LiDAR segmentation is crucial for autonomous driving perception. Recent trends favor point- or voxel-based methods as they often yield better performance than the traditional range view representation. In this work, we unveil several key factors in building powerful range view models. We observe that the "many-to-one" mapping, semantic incoherence, and shape deformation are possible impediments against effective learning from range view projections. We present RangeFormer -- a full-cycle framework comprising novel designs across network architecture, data augmentation, and post-processing -- that better handles the learning and processing of LiDAR point clouds from the range view. We further introduce a Scalable Training from Range view (STR) strategy that trains on arbitrary low-resolution 2D range images, while still maintaining satisfactory 3D segmentation accuracy. We show that, for the first time, a range view method is able to surpass the point, voxel, and multi-view fusion counterparts in the competing LiDAR semantic and panoptic segmentation benchmarks, i.e., SemanticKITTI, nuScenes, and ScribbleKITTI.

Zhaoyang Xia, Youquan Liu, Xin Li et al.

Training deep models for semantic scene completion (SSC) is challenging due to the sparse and incomplete input, a large quantity of objects of diverse scales as well as the inherent label noise for moving objects. To address the above-mentioned problems, we propose the following three solutions: 1) Redesigning the completion sub-network. We design a novel completion sub-network, which consists of several Multi-Path Blocks (MPBs) to aggregate multi-scale features and is free from the lossy downsampling operations. 2) Distilling rich knowledge from the multi-frame model. We design a novel knowledge distillation objective, dubbed Dense-to-Sparse Knowledge Distillation (DSKD). It transfers the dense, relation-based semantic knowledge from the multi-frame teacher to the single-frame student, significantly improving the representation learning of the single-frame model. 3) Completion label rectification. We propose a simple yet effective label rectification strategy, which uses off-the-shelf panoptic segmentation labels to remove the traces of dynamic objects in completion labels, greatly improving the performance of deep models especially for those moving objects. Extensive experiments are conducted in two public SSC benchmarks, i.e., SemanticKITTI and SemanticPOSS. Our SCPNet ranks 1st on SemanticKITTI semantic scene completion challenge and surpasses the competitive S3CNet by 7.2 mIoU. SCPNet also outperforms previous completion algorithms on the SemanticPOSS dataset. Besides, our method also achieves competitive results on SemanticKITTI semantic segmentation tasks, showing that knowledge learned in the scene completion is beneficial to the segmentation task.

Guodong Xu, Yuenan Hou, Ziwei Liu et al.

StyleGAN family is one of the most popular Generative Adversarial Networks (GANs) for unconditional generation. Despite its impressive performance, its high demand on storage and computation impedes their deployment on resource-constrained devices. This paper provides a comprehensive study of distilling from the popular StyleGAN-like architecture. Our key insight is that the main challenge of StyleGAN distillation lies in the output discrepancy issue, where the teacher and student model yield different outputs given the same input latent code. Standard knowledge distillation losses typically fail under this heterogeneous distillation scenario. We conduct thorough analysis about the reasons and effects of this discrepancy issue, and identify that the mapping network plays a vital role in determining semantic information of generated images. Based on this finding, we propose a novel initialization strategy for the student model, which can ensure the output consistency to the maximum extent. To further enhance the semantic consistency between the teacher and student model, we present a latent-direction-based distillation loss that preserves the semantic relations in latent space. Extensive experiments demonstrate the effectiveness of our approach in distilling StyleGAN2 and StyleGAN3, outperforming existing GAN distillation methods by a large margin.

Xiaopei Wu, Liang Peng, Liang Xie et al.

Semi-supervised learning aims to leverage numerous unlabeled data to improve the model performance. Current semi-supervised 3D object detection methods typically use a teacher to generate pseudo labels for a student, and the quality of the pseudo labels is essential for the final performance. In this paper, we propose PatchTeacher, which focuses on partial scene 3D object detection to provide high-quality pseudo labels for the student. Specifically, we divide a complete scene into a series of patches and feed them to our PatchTeacher sequentially. PatchTeacher leverages the low memory consumption advantage of partial scene detection to process point clouds with a high-resolution voxelization, which can minimize the information loss of quantization and extract more fine-grained features. However, it is non-trivial to train a detector on fractions of the scene. Therefore, we introduce three key techniques, i.e., Patch Normalizer, Quadrant Align, and Fovea Selection, to improve the performance of PatchTeacher. Moreover, we devise PillarMix, a strong data augmentation strategy that mixes truncated pillars from different LiDAR scans to generate diverse training samples and thus help the model learn more general representation. Extensive experiments conducted on Waymo and ONCE datasets verify the effectiveness and superiority of our method and we achieve new state-of-the-art results, surpassing existing methods by a large margin. Codes are available at https://github.com/LittlePey/PTPM.

Zijia Chen, Yuenan Hou, Xinhua Jiang et al.

Robotic manipulation requires the effective integration of heterogeneous inputs, including visual observations, language instructions, and trajectory representations, to generate accurate actions. Existing transformer-based policies typically process these heterogeneous modalities within a shared parameter space, which often leads to modality interference and inefficient representation learning, especially in data-scarce scenarios. While Mixture-of-Experts (MoE) offers a scalable solution through expert specialization, conventional routing mechanisms are often sensitive to such cross-modal representation discrepancies, resulting in unstable expert assignment and expert collapse. In this work, we propose MATE (Multi-ModAl TrajEctory Policies), a novel trajectory prediction framework built upon MoE. Specifically, we introduce a Multi-Modal MoE architecture to achieve fine-grained sub-token feature decoupling, and design a cross-modal cosine router for stable and scale-invariant expert assignment across heterogeneous modalities. We further employ temperature-controlled routing and stochastic noise injection to improve expert balance and prevent premature routing collapse under scarce demonstrations. Experiments on the LIBERO benchmark show that our MATE consistently outperforms prior work under data scarcity. It achieves a 4.75% improvement in average success rate over the trajectory-guided counterpart. Real-world experiments on robotic ping-pong also suggest that the predicted trajectories can provide useful guidance for downstream robotic execution, further indicating the practical feasibility of our algorithm.

Yiteng Xu, Peishan Cong, Yichen Yao et al.

Human-centric scene understanding is significant for real-world applications, but it is extremely challenging due to the existence of diverse human poses and actions, complex human-environment interactions, severe occlusions in crowds, etc. In this paper, we present a large-scale multi-modal dataset for human-centric scene understanding, dubbed HuCenLife, which is collected in diverse daily-life scenarios with rich and fine-grained annotations. Our HuCenLife can benefit many 3D perception tasks, such as segmentation, detection, action recognition, etc., and we also provide benchmarks for these tasks to facilitate related research. In addition, we design novel modules for LiDAR-based segmentation and action recognition, which are more applicable for large-scale human-centric scenarios and achieve state-of-the-art performance.

Yuhang Lu, Qi Jiang, Runnan Chen et al.

Zero-shot point cloud segmentation aims to make deep models capable of recognizing novel objects in point cloud that are unseen in the training phase. Recent trends favor the pipeline which transfers knowledge from seen classes with labels to unseen classes without labels. They typically align visual features with semantic features obtained from word embedding by the supervision of seen classes' annotations. However, point cloud contains limited information to fully match with semantic features. In fact, the rich appearance information of images is a natural complement to the textureless point cloud, which is not well explored in previous literature. Motivated by this, we propose a novel multi-modal zero-shot learning method to better utilize the complementary information of point clouds and images for more accurate visual-semantic alignment. Extensive experiments are performed in two popular benchmarks, i.e., SemanticKITTI and nuScenes, and our method outperforms current SOTA methods with 52% and 49% improvement on average for unseen class mIoU, respectively.

Yu Li, Yuenan Hou, Yingmei Wei et al.

Large Language Models (LLMs) remain highly vulnerable to diverse attacks, particularly in black-box settings where the internals of target models are inaccessible. Existing black-box defenses typically rely on pre-defined filtering heuristics, which often fail to generalize to unseen attack types and target model architectures. We introduce EvoDefense, an experience-guided co-evolving black-box defense paradigm. EvoDefense employs a guard LLM to detect malicious queries and an experience memory module to accumulate defense knowledge from previous interactions. At the core of EvoDefense is a continuous attack-defense evolution loop, where an attack generator and the guard model iteratively refine their attack strategies and defense policies through experience-guided optimization. This design enables EvoDefense to generalize across unseen attacks and target models without retraining. Experiments on HarmBench, AdvBench, and AlpacaEval show that EvoDefense achieves consistently strong defense performance across seven popular models and five representative LLM attacks, while preserving competitive general capabilities. On HarmBench, EvoDefense reduces the attack success rate (ASR) of AutoDAN-turbo on Gemini-3-flash and LLaMA-3-8B-Instruct from 29.4% and 43.4% to 8.4% and 6.2%, respectively.

Xin Li, Botian Shi, Yuenan Hou et al.

Multi-modal 3D object detection has been an active research topic in autonomous driving. Nevertheless, it is non-trivial to explore the cross-modal feature fusion between sparse 3D points and dense 2D pixels. Recent approaches either fuse the image features with the point cloud features that are projected onto the 2D image plane or combine the sparse point cloud with dense image pixels. These fusion approaches often suffer from severe information loss, thus causing sub-optimal performance. To address these problems, we construct the homogeneous structure between the point cloud and images to avoid projective information loss by transforming the camera features into the LiDAR 3D space. In this paper, we propose a homogeneous multi-modal feature fusion and interaction method (HMFI) for 3D object detection. Specifically, we first design an image voxel lifter module (IVLM) to lift 2D image features into the 3D space and generate homogeneous image voxel features. Then, we fuse the voxelized point cloud features with the image features from different regions by introducing the self-attention based query fusion mechanism (QFM). Next, we propose a voxel feature interaction module (VFIM) to enforce the consistency of semantic information from identical objects in the homogeneous point cloud and image voxel representations, which can provide object-level alignment guidance for cross-modal feature fusion and strengthen the discriminative ability in complex backgrounds. We conduct extensive experiments on the KITTI and Waymo Open Dataset, and the proposed HMFI achieves better performance compared with the state-of-the-art multi-modal methods. Particularly, for the 3D detection of cyclist on the KITTI benchmark, HMFI surpasses all the published algorithms by a large margin.

Heng Li, Yuenan Hou, Xiaohan Xing et al.

Training deep learning models for semantic occupancy prediction is challenging due to factors such as a large number of occupancy cells, severe occlusion, limited visual cues, complicated driving scenarios, etc. Recent methods often adopt transformer-based architectures given their strong capability in learning input-conditioned weights and long-range relationships. However, transformer-based networks are notorious for their quadratic computation complexity, seriously undermining their efficacy and deployment in semantic occupancy prediction. Inspired by the global modeling and linear computation complexity of the Mamba architecture, we present the first Mamba-based network for semantic occupancy prediction, termed OccMamba. Specifically, we first design the hierarchical Mamba module and local context processor to better aggregate global and local contextual information, respectively. Besides, to relieve the inherent domain gap between the linguistic and 3D domains, we present a simple yet effective 3D-to-1D reordering scheme, i.e., height-prioritized 2D Hilbert expansion. It can maximally retain the spatial structure of 3D voxels as well as facilitate the processing of Mamba blocks. Endowed with the aforementioned designs, our OccMamba is capable of directly and efficiently processing large volumes of dense scene grids, achieving state-of-the-art performance across three prevalent occupancy prediction benchmarks, including OpenOccupancy, SemanticKITTI, and SemanticPOSS. Notably, on OpenOccupancy, our OccMamba outperforms the previous state-of-the-art Co-Occ by 5.1% IoU and 4.3% mIoU, respectively. Our implementation is open-sourced and available at: https://github.com/USTCLH/OccMamba.

Xiao Zheng, Xiaoshui Huang, Guofeng Mei et al.

Pre-training a model and then fine-tuning it on downstream tasks has demonstrated significant success in the 2D image and NLP domains. However, due to the unordered and non-uniform density characteristics of point clouds, it is non-trivial to explore the prior knowledge of point clouds and pre-train a point cloud backbone. In this paper, we propose a novel pre-training method called Point cloud Diffusion pre-training (PointDif). We consider the point cloud pre-training task as a conditional point-to-point generation problem and introduce a conditional point generator. This generator aggregates the features extracted by the backbone and employs them as the condition to guide the point-to-point recovery from the noisy point cloud, thereby assisting the backbone in capturing both local and global geometric priors as well as the global point density distribution of the object. We also present a recurrent uniform sampling optimization strategy, which enables the model to uniformly recover from various noise levels and learn from balanced supervision. Our PointDif achieves substantial improvement across various real-world datasets for diverse downstream tasks such as classification, segmentation and detection. Specifically, PointDif attains 70.0% mIoU on S3DIS Area 5 for the segmentation task and achieves an average improvement of 2.4% on ScanObjectNN for the classification task compared to TAP. Furthermore, our pre-training framework can be flexibly applied to diverse point cloud backbones and bring considerable gains.

Jian Liu, Xiaoshui Huang, Tianyu Huang et al.

Recent years have witnessed remarkable advances in artificial intelligence generated content(AIGC), with diverse input modalities, e.g., text, image, video, audio and 3D. The 3D is the most close visual modality to real-world 3D environment and carries enormous knowledge. The 3D content generation shows both academic and practical values while also presenting formidable technical challenges. This review aims to consolidate developments within the burgeoning domain of 3D content generation. Specifically, a new taxonomy is proposed that categorizes existing approaches into three types: 3D native generative methods, 2D prior-based 3D generative methods, and hybrid 3D generative methods. The survey covers approximately 60 papers spanning the major techniques. Besides, we discuss limitations of current 3D content generation techniques, and point out open challenges as well as promising directions for future work. Accompanied with this survey, we have established a project website where the resources on 3D content generation research are provided. The project page is available at https://github.com/hitcslj/Awesome-AIGC-3D.

Weijie Li, Wei Yang, Yuenan Hou et al.

Despite the remarkable progress in synthetic aperture radar automatic target recognition (SAR ATR), recent efforts have concentrated on detecting and classifying a specific category, e.g., vehicles, ships, airplanes, or buildings. One of the fundamental limitations of the top-performing SAR ATR methods is that the learning paradigm is supervised, task-specific, limited-category, closed-world learning, which depends on massive amounts of accurately annotated samples that are expensively labeled by expert SAR analysts and have limited generalization capability and scalability. In this work, we make the first attempt towards building a foundation model for SAR ATR, termed SARATR-X. SARATR-X learns generalizable representations via self-supervised learning (SSL) and provides a cornerstone for label-efficient model adaptation to generic SAR target detection and classification tasks. Specifically, SARATR-X is trained on 0.18 M unlabelled SAR target samples, which are curated by combining contemporary benchmarks and constitute the largest publicly available dataset till now. Considering the characteristics of SAR images, a backbone tailored for SAR ATR is carefully designed, and a two-step SSL method endowed with multi-scale gradient features was applied to ensure the feature diversity and model scalability of SARATR-X. The capabilities of SARATR-X are evaluated on classification under few-shot and robustness settings and detection across various categories and scenes, and impressive performance is achieved, often competitive with or even superior to prior fully supervised, semi-supervised, or self-supervised algorithms. Our SARATR-X and the curated dataset are released at https://github.com/waterdisappear/SARATR-X to foster research into foundation models for SAR image interpretation.

Chenxi Huang, Yuenan Hou, Weicai Ye et al.

NeRF-Det has achieved impressive performance in indoor multi-view 3D detection by innovatively utilizing NeRF to enhance representation learning. Despite its notable performance, we uncover three decisive shortcomings in its current design, including semantic ambiguity, inappropriate sampling, and insufficient utilization of depth supervision. To combat the aforementioned problems, we present three corresponding solutions: 1) Semantic Enhancement. We project the freely available 3D segmentation annotations onto the 2D plane and leverage the corresponding 2D semantic maps as the supervision signal, significantly enhancing the semantic awareness of multi-view detectors. 2) Perspective-aware Sampling. Instead of employing the uniform sampling strategy, we put forward the perspective-aware sampling policy that samples densely near the camera while sparsely in the distance, more effectively collecting the valuable geometric clues. 3)Ordinal Residual Depth Supervision. As opposed to directly regressing the depth values that are difficult to optimize, we divide the depth range of each scene into a fixed number of ordinal bins and reformulate the depth prediction as the combination of the classification of depth bins as well as the regression of the residual depth values, thereby benefiting the depth learning process. The resulting algorithm, NeRF-Det++, has exhibited appealing performance in the ScanNetV2 and ARKITScenes datasets. Notably, in ScanNetV2, NeRF-Det++ outperforms the competitive NeRF-Det by +1.9% in mAP@0.25 and +3.5% in mAP@0.50$. The code will be publicly at https://github.com/mrsempress/NeRF-Detplusplus.

Linfeng Dong, Yuchen Yang, Hao Wu et al.

We introduce RacketVision, a novel dataset and benchmark for advancing computer vision in sports analytics, covering table tennis, tennis, and badminton. The dataset is the first to provide large-scale, fine-grained annotations for racket pose alongside traditional ball positions, enabling research into complex human-object interactions. It is designed to tackle three interconnected tasks: fine-grained ball tracking, articulated racket pose estimation, and predictive ball trajectory forecasting. Our evaluation of established baselines reveals a critical insight for multi-modal fusion: while naively concatenating racket pose features degrades performance, a CrossAttention mechanism is essential to unlock their value, leading to trajectory prediction results that surpass strong unimodal baselines. RacketVision provides a versatile resource and a strong starting point for future research in dynamic object tracking, conditional motion forecasting, and multimodal analysis in sports. Project page at https://github.com/OrcustD/RacketVision

Yuenan Hou, Zheng Ma, Chunxiao Liu et al.

Channel pruning is broadly recognized as an effective approach to obtain a small compact model through eliminating unimportant channels from a large cumbersome network. Contemporary methods typically perform iterative pruning procedure from the original over-parameterized model, which is both tedious and expensive especially when the pruning is aggressive. In this paper, we propose a simple yet effective channel pruning technique, termed network Pruning via rEsource rEalLocation (PEEL), to quickly produce a desired slim model with negligible cost. Specifically, PEEL first constructs a predefined backbone and then conducts resource reallocation on it to shift parameters from less informative layers to more important layers in one round, thus amplifying the positive effect of these informative layers. To demonstrate the effectiveness of PEEL , we perform extensive experiments on ImageNet with ResNet-18, ResNet-50, MobileNetV2, MobileNetV3-small and EfficientNet-B0. Experimental results show that structures uncovered by PEEL exhibit competitive performance with state-of-the-art pruning algorithms under various pruning settings. Our code is available at https://github.com/cardwing/Codes-for-PEEL.

Yuenan Hou, Zheng Ma, Chunxiao Liu et al.

We study the problem of distilling knowledge from a large deep teacher network to a much smaller student network for the task of road marking segmentation. In this work, we explore a novel knowledge distillation (KD) approach that can transfer 'knowledge' on scene structure more effectively from a teacher to a student model. Our method is known as Inter-Region Affinity KD (IntRA-KD). It decomposes a given road scene image into different regions and represents each region as a node in a graph. An inter-region affinity graph is then formed by establishing pairwise relationships between nodes based on their similarity in feature distribution. To learn structural knowledge from the teacher network, the student is required to match the graph generated by the teacher. The proposed method shows promising results on three large-scale road marking segmentation benchmarks, i.e., ApolloScape, CULane and LLAMAS, by taking various lightweight models as students and ResNet-101 as the teacher. IntRA-KD consistently brings higher performance gains on all lightweight models, compared to previous distillation methods. Our code is available at https://github.com/cardwing/Codes-for-IntRA-KD.

Yuenan Hou, Zheng Ma, Chunxiao Liu et al.

Training deep models for lane detection is challenging due to the very subtle and sparse supervisory signals inherent in lane annotations. Without learning from much richer context, these models often fail in challenging scenarios, e.g., severe occlusion, ambiguous lanes, and poor lighting conditions. In this paper, we present a novel knowledge distillation approach, i.e., Self Attention Distillation (SAD), which allows a model to learn from itself and gains substantial improvement without any additional supervision or labels. Specifically, we observe that attention maps extracted from a model trained to a reasonable level would encode rich contextual information. The valuable contextual information can be used as a form of 'free' supervision for further representation learning through performing topdown and layer-wise attention distillation within the network itself. SAD can be easily incorporated in any feedforward convolutional neural networks (CNN) and does not increase the inference time. We validate SAD on three popular lane detection benchmarks (TuSimple, CULane and BDD100K) using lightweight models such as ENet, ResNet-18 and ResNet-34. The lightest model, ENet-SAD, performs comparatively or even surpasses existing algorithms. Notably, ENet-SAD has 20 x fewer parameters and runs 10 x faster compared to the state-of-the-art SCNN, while still achieving compelling performance in all benchmarks. Our code is available at https://github.com/cardwing/Codes-for-Lane-Detection.

Dingning Liu, Xiaoshui Huang, Yuenan Hou et al.

In this paper, we introduce Uni3D-LLM, a unified framework that leverages a Large Language Model (LLM) to integrate tasks of 3D perception, generation, and editing within point cloud scenes. This framework empowers users to effortlessly generate and modify objects at specified locations within a scene, guided by the versatility of natural language descriptions. Uni3D-LLM harnesses the expressive power of natural language to allow for precise command over the generation and editing of 3D objects, thereby significantly enhancing operational flexibility and controllability. By mapping point cloud into the unified representation space, Uni3D-LLM achieves cross-application functionality, enabling the seamless execution of a wide array of tasks, ranging from the accurate instantiation of 3D objects to the diverse requirements of interactive design. Through a comprehensive suite of rigorous experiments, the efficacy of Uni3D-LLM in the comprehension, generation, and editing of point cloud has been validated. Additionally, we have assessed the impact of integrating a point cloud perception module on the generation and editing processes, confirming the substantial potential of our approach for practical applications.

Yu Li, Yuenan Hou, Yingmei Wei et al.

Multi-modal 3D understanding is a fundamental task in computer vision. Previous multi-modal fusion methods typically employ a single, dense fusion network, struggling to handle the significant heterogeneity and complexity across modalities, leading to suboptimal performance. In this paper, we propose MoE3D, which integrates Mixture of Experts (MoE) into the multi-modal learning framework. The core is that we deploy a set of specialized "expert" networks, each adept at processing a specific modality or a mode of cross-modal interaction. Specifically, the MoE-based transformer is designed to better utilize the complementary information hidden in the visual features. Information aggregation module is put forward to further enhance the fusion performance. Top-1 gating is employed to make one expert process features with expert groups, ensuring high efficiency. We further propose a progressive pre-training strategy to better leverage the semantic and 2D prior, thus equipping the network with good initialization. Our MoE3D achieves competitive performance across four prevalent 3D understanding tasks. Notably, our MoE3D surpasses the top-performing counterpart by 6.1 mIoU on Multi3DRefer.

Shanghai AI Lab, Yicheng Bao, Guanxu Chen et al.

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

Zeren Chen, Yuenan Hou, Yulin Chen et al.

In this paper, we introduce the HexPlane representation for 3D semantic scene understanding. Specifically, we first design the View Projection Module (VPM) to project the 3D point cloud into six planes to maximally retain the original spatial information. Features of six planes are extracted by the 2D encoder and sent to the HexPlane Association Module (HAM) to adaptively fuse the most informative information for each point. The fused point features are further fed to the task head to yield the ultimate predictions. Compared to the popular point and voxel representation, the HexPlane representation is efficient and can utilize highly optimized 2D operations to process sparse and unordered 3D point clouds. It can also leverage off-the-shelf 2D models, network weights, and training recipes to achieve accurate scene understanding in 3D space. On ScanNet and SemanticKITTI benchmarks, our algorithm, dubbed HexNet3D, achieves competitive performance with previous algorithms. In particular, on the ScanNet 3D segmentation task, our method obtains 77.0 mIoU on the validation set, surpassing Point Transformer V2 by 1.6 mIoU. We also observe encouraging results in indoor 3D detection tasks. Note that our method can be seamlessly integrated into existing voxel-based, point-based, and range-based approaches and brings considerable gains without bells and whistles. The codes will be available upon publication.

Xulin Li, Yan Lu, Bin Liu et al.

Clothes-invariant feature extraction is critical to the clothes-changing person re-identification (CC-ReID). It can provide discriminative identity features and eliminate the negative effects caused by the confounder--clothing changes. But we argue that there exists a strong spurious correlation between clothes and human identity, that restricts the common likelihood-based ReID method P(Y|X) to extract clothes-irrelevant features. In this paper, we propose a new Causal Clothes-Invariant Learning (CCIL) method to achieve clothes-invariant feature learning by modeling causal intervention P(Y|do(X)). This new causality-based model is inherently invariant to the confounder in the causal view, which can achieve the clothes-invariant features and avoid the barrier faced by the likelihood-based methods. Extensive experiments on three CC-ReID benchmarks, including PRCC, LTCC, and VC-Clothes, demonstrate the effectiveness of our approach, which achieves a new state of the art.

Xiaohan Xing, Yuenan Hou, Hang Li et al.

The amount of medical images for training deep classification models is typically very scarce, making these deep models prone to overfit the training data. Studies showed that knowledge distillation (KD), especially the mean-teacher framework which is more robust to perturbations, can help mitigate the over-fitting effect. However, directly transferring KD from computer vision to medical image classification yields inferior performance as medical images suffer from higher intra-class variance and class imbalance. To address these issues, we propose a novel Categorical Relation-preserving Contrastive Knowledge Distillation (CRCKD) algorithm, which takes the commonly used mean-teacher model as the supervisor. Specifically, we propose a novel Class-guided Contrastive Distillation (CCD) module to pull closer positive image pairs from the same class in the teacher and student models, while pushing apart negative image pairs from different classes. With this regularization, the feature distribution of the student model shows higher intra-class similarity and inter-class variance. Besides, we propose a Categorical Relation Preserving (CRP) loss to distill the teacher's relational knowledge in a robust and class-balanced manner. With the contribution of the CCD and CRP, our CRCKD algorithm can distill the relational knowledge more comprehensively. Extensive experiments on the HAM10000 and APTOS datasets demonstrate the superiority of the proposed CRCKD method.

Yuenan Hou

Lane detection is an important yet challenging task in autonomous driving, which is affected by many factors, e.g., light conditions, occlusions caused by other vehicles, irrelevant markings on the road and the inherent long and thin property of lanes. Conventional methods typically treat lane detection as a semantic segmentation task, which assigns a class label to each pixel of the image. This formulation heavily depends on the assumption that the number of lanes is pre-defined and fixed and no lane changing occurs, which does not always hold. To make the lane detection model applicable to an arbitrary number of lanes and lane changing scenarios, we adopt an instance segmentation approach, which first differentiates lanes and background and then classify each lane pixel into each lane instance. Besides, a multi-task learning paradigm is utilized to better exploit the structural information and the feature pyramid architecture is used to detect extremely thin lanes. Three popular lane detection benchmarks, i.e., TuSimple, CULane and BDD100K, are used to validate the effectiveness of our proposed algorithm.

Yuenan Hou, Zheng Ma, Chunxiao Liu et al.

The training of many existing end-to-end steering angle prediction models heavily relies on steering angles as the supervisory signal. Without learning from much richer contexts, these methods are susceptible to the presence of sharp road curves, challenging traffic conditions, strong shadows, and severe lighting changes. In this paper, we considerably improve the accuracy and robustness of predictions through heterogeneous auxiliary networks feature mimicking, a new and effective training method that provides us with much richer contextual signals apart from steering direction. Specifically, we train our steering angle predictive model by distilling multi-layer knowledge from multiple heterogeneous auxiliary networks that perform related but different tasks, e.g., image segmentation or optical flow estimation. As opposed to multi-task learning, our method does not require expensive annotations of related tasks on the target set. This is made possible by applying contemporary off-the-shelf networks on the target set and mimicking their features in different layers after transformation. The auxiliary networks are discarded after training without affecting the runtime efficiency of our model. Our approach achieves a new state-of-the-art on Udacity and Comma.ai, outperforming the previous best by a large margin of 12.8% and 52.1%, respectively. Encouraging results are also shown on Berkeley Deep Drive (BDD) dataset.