Zeming Li

Yanwei Li, Yilun Chen, Xiaojuan Qi et al.

In this work, we present a unified framework for multi-modality 3D object detection, named UVTR. The proposed method aims to unify multi-modality representations in the voxel space for accurate and robust single- or cross-modality 3D detection. To this end, the modality-specific space is first designed to represent different inputs in the voxel feature space. Different from previous work, our approach preserves the voxel space without height compression to alleviate semantic ambiguity and enable spatial connections. To make full use of the inputs from different sensors, the cross-modality interaction is then proposed, including knowledge transfer and modality fusion. In this way, geometry-aware expressions in point clouds and context-rich features in images are well utilized for better performance and robustness. The transformer decoder is applied to efficiently sample features from the unified space with learnable positions, which facilitates object-level interactions. In general, UVTR presents an early attempt to represent different modalities in a unified framework. It surpasses previous work in single- or multi-modality entries. The proposed method achieves leading performance in the nuScenes test set for both object detection and the following object tracking task. Code is made publicly available at https://github.com/dvlab-research/UVTR.

Yanwei Li, Xiaojuan Qi, Yukang Chen et al.

In this work, we present a conceptually simple yet effective framework for cross-modality 3D object detection, named voxel field fusion. The proposed approach aims to maintain cross-modality consistency by representing and fusing augmented image features as a ray in the voxel field. To this end, the learnable sampler is first designed to sample vital features from the image plane that are projected to the voxel grid in a point-to-ray manner, which maintains the consistency in feature representation with spatial context. In addition, ray-wise fusion is conducted to fuse features with the supplemental context in the constructed voxel field. We further develop mixed augmentor to align feature-variant transformations, which bridges the modality gap in data augmentation. The proposed framework is demonstrated to achieve consistent gains in various benchmarks and outperforms previous fusion-based methods on KITTI and nuScenes datasets. Code is made available at https://github.com/dvlab-research/VFF.

Shaoru Wang, Jin Gao, Zeming Li et al.

Self-supervised learning on large-scale Vision Transformers (ViTs) as pre-training methods has achieved promising downstream performance. Yet, how much these pre-training paradigms promote lightweight ViTs' performance is considerably less studied. In this work, we develop and benchmark several self-supervised pre-training methods on image classification tasks and some downstream dense prediction tasks. We surprisingly find that if proper pre-training is adopted, even vanilla lightweight ViTs show comparable performance to previous SOTA networks with delicate architecture design. It breaks the recently popular conception that vanilla ViTs are not suitable for vision tasks in lightweight regimes. We also point out some defects of such pre-training, e.g., failing to benefit from large-scale pre-training data and showing inferior performance on data-insufficient downstream tasks. Furthermore, we analyze and clearly show the effect of such pre-training by analyzing the properties of the layer representation and attention maps for related models. Finally, based on the above analyses, a distillation strategy during pre-training is developed, which leads to further downstream performance improvement for MAE-based pre-training. Code is available at https://github.com/wangsr126/mae-lite.

Yinhao Li, Han Bao, Zheng Ge et al.

Bounded by the inherent ambiguity of depth perception, contemporary camera-based 3D object detection methods fall into the performance bottleneck. Intuitively, leveraging temporal multi-view stereo (MVS) technology is the natural knowledge for tackling this ambiguity. However, traditional attempts of MVS are flawed in two aspects when applying to 3D object detection scenes: 1) The affinity measurement among all views suffers expensive computation cost; 2) It is difficult to deal with outdoor scenarios where objects are often mobile. To this end, we introduce an effective temporal stereo method to dynamically select the scale of matching candidates, enable to significantly reduce computation overhead. Going one step further, we design an iterative algorithm to update more valuable candidates, making it adaptive to moving candidates. We instantiate our proposed method to multi-view 3D detector, namely BEVStereo. BEVStereo achieves the new state-of-the-art performance (i.e., 52.5% mAP and 61.0% NDS) on the camera-only track of nuScenes dataset. Meanwhile, extensive experiments reflect our method can deal with complex outdoor scenarios better than contemporary MVS approaches. Codes have been released at https://github.com/Megvii-BaseDetection/BEVStereo.

Zhisheng Zhong, Jiequan Cui, Zeming Li et al.

Deep neural networks perform poorly on heavily class-imbalanced datasets. Given the promising performance of contrastive learning, we propose Rebalanced Siamese Contrastive Mining (ResCom) to tackle imbalanced recognition. Based on the mathematical analysis and simulation results, we claim that supervised contrastive learning suffers a dual class-imbalance problem at both the original batch and Siamese batch levels, which is more serious than long-tailed classification learning. In this paper, at the original batch level, we introduce a class-balanced supervised contrastive loss to assign adaptive weights for different classes. At the Siamese batch level, we present a class-balanced queue, which maintains the same number of keys for all classes. Furthermore, we note that the imbalanced contrastive loss gradient with respect to the contrastive logits can be decoupled into the positives and negatives, and easy positives and easy negatives will make the contrastive gradient vanish. We propose supervised hard positive and negative pairs mining to pick up informative pairs for contrastive computation and improve representation learning. Finally, to approximately maximize the mutual information between the two views, we propose Siamese Balanced Softmax and joint it with the contrastive loss for one-stage training. Extensive experiments demonstrate that ResCom outperforms the previous methods by large margins on multiple long-tailed recognition benchmarks. Our code and models are made publicly available at: https://github.com/dvlab-research/ResCom.

Jinrong Yang, Songtao Liu, Zeming Li et al.

Autonomous driving requires the model to perceive the environment and (re)act within a low latency for safety. While past works ignore the inevitable changes in the environment after processing, streaming perception is proposed to jointly evaluate the latency and accuracy into a single metric for video online perception. In this paper, instead of searching trade-offs between accuracy and speed like previous works, we point out that endowing real-time models with the ability to predict the future is the key to dealing with this problem. We build a simple and effective framework for streaming perception. It equips a novel DualFlow Perception module (DFP), which includes dynamic and static flows to capture the moving trend and basic detection feature for streaming prediction. Further, we introduce a Trend-Aware Loss (TAL) combined with a trend factor to generate adaptive weights for objects with different moving speeds. Our simple method achieves competitive performance on Argoverse-HD dataset and improves the AP by 4.9% compared to the strong baseline, validating its effectiveness. Our code will be made available at https://github.com/yancie-yjr/StreamYOLO.

Lin Song, Songyang Zhang, Songtao Liu et al.

Transformers, the de-facto standard for language modeling, have been recently applied for vision tasks. This paper introduces sparse queries for vision transformers to exploit the intrinsic spatial redundancy of natural images and save computational costs. Specifically, we propose a Dynamic Grained Encoder for vision transformers, which can adaptively assign a suitable number of queries to each spatial region. Thus it achieves a fine-grained representation in discriminative regions while keeping high efficiency. Besides, the dynamic grained encoder is compatible with most vision transformer frameworks. Without bells and whistles, our encoder allows the state-of-the-art vision transformers to reduce computational complexity by 40%-60% while maintaining comparable performance on image classification. Extensive experiments on object detection and segmentation further demonstrate the generalizability of our approach. Code is available at https://github.com/StevenGrove/vtpack.

Jinrong Yang, Lin Song, Songtao Liu et al.

Many point-based 3D detectors adopt point-feature sampling strategies to drop some points for efficient inference. These strategies are typically based on fixed and handcrafted rules, making it difficult to handle complicated scenes. Different from them, we propose a Dynamic Ball Query (DBQ) network to adaptively select a subset of input points according to the input features, and assign the feature transform with a suitable receptive field for each selected point. It can be embedded into some state-of-the-art 3D detectors and trained in an end-to-end manner, which significantly reduces the computational cost. Extensive experiments demonstrate that our method can increase the inference speed by 30%-100% on KITTI, Waymo, and ONCE datasets. Specifically, the inference speed of our detector can reach 162 FPS on KITTI scene, and 30 FPS on Waymo and ONCE scenes without performance degradation. Due to skipping the redundant points, some evaluation metrics show significant improvements. Codes will be released at https://github.com/yancie-yjr/DBQ-SSD.

Yinhao Li, Zheng Ge, Guanyi Yu et al.

In this research, we propose a new 3D object detector with a trustworthy depth estimation, dubbed BEVDepth, for camera-based Bird's-Eye-View (BEV) 3D object detection. Our work is based on a key observation -- depth estimation in recent approaches is surprisingly inadequate given the fact that depth is essential to camera 3D detection. Our BEVDepth resolves this by leveraging explicit depth supervision. A camera-awareness depth estimation module is also introduced to facilitate the depth predicting capability. Besides, we design a novel Depth Refinement Module to counter the side effects carried by imprecise feature unprojection. Aided by customized Efficient Voxel Pooling and multi-frame mechanism, BEVDepth achieves the new state-of-the-art 60.9% NDS on the challenging nuScenes test set while maintaining high efficiency. For the first time, the NDS score of a camera model reaches 60%.

Hongyu Zhou, Zheng Ge, Songtao Liu et al.

To date, the most powerful semi-supervised object detectors (SS-OD) are based on pseudo-boxes, which need a sequence of post-processing with fine-tuned hyper-parameters. In this work, we propose replacing the sparse pseudo-boxes with the dense prediction as a united and straightforward form of pseudo-label. Compared to the pseudo-boxes, our Dense Pseudo-Label (DPL) does not involve any post-processing method, thus retaining richer information. We also introduce a region selection technique to highlight the key information while suppressing the noise carried by dense labels. We name our proposed SS-OD algorithm that leverages the DPL as Dense Teacher. On COCO and VOC, Dense Teacher shows superior performance under various settings compared with the pseudo-box-based methods.

Zengran Wang, Chen Min, Zheng Ge et al.

Learning accurate depth is essential to multi-view 3D object detection. Recent approaches mainly learn depth from monocular images, which confront inherent difficulties due to the ill-posed nature of monocular depth learning. Instead of using a sole monocular depth method, in this work, we propose a novel Surround-view Temporal Stereo (STS) technique that leverages the geometry correspondence between frames across time to facilitate accurate depth learning. Specifically, we regard the field of views from all cameras around the ego vehicle as a unified view, namely surroundview, and conduct temporal stereo matching on it. The resulting geometrical correspondence between different frames from STS is utilized and combined with the monocular depth to yield final depth prediction. Comprehensive experiments on nuScenes show that STS greatly boosts 3D detection ability, notably for medium and long distance objects. On BEVDepth with ResNet-50 backbone, STS improves mAP and NDS by 2.6% and 1.4%, respectively. Consistent improvements are observed when using a larger backbone and a larger image resolution, demonstrating its effectiveness

Hongyu Zhou, Zheng Ge, Zeming Li et al.

This paper proposes an efficient multi-camera to Bird's-Eye-View (BEV) view transformation method for 3D perception, dubbed MatrixVT. Existing view transformers either suffer from poor transformation efficiency or rely on device-specific operators, hindering the broad application of BEV models. In contrast, our method generates BEV features efficiently with only convolutions and matrix multiplications (MatMul). Specifically, we propose describing the BEV feature as the MatMul of image feature and a sparse Feature Transporting Matrix (FTM). A Prime Extraction module is then introduced to compress the dimension of image features and reduce FTM's sparsity. Moreover, we propose the Ring \& Ray Decomposition to replace the FTM with two matrices and reformulate our pipeline to reduce calculation further. Compared to existing methods, MatrixVT enjoys a faster speed and less memory footprint while remaining deploy-friendly. Extensive experiments on the nuScenes benchmark demonstrate that our method is highly efficient but obtains results on par with the SOTA method in object detection and map segmentation tasks

En Yu, Songtao Liu, Zhuoling Li et al.

Although existing multi-object tracking (MOT) algorithms have obtained competitive performance on various benchmarks, almost all of them train and validate models on the same domain. The domain generalization problem of MOT is hardly studied. To bridge this gap, we first draw the observation that the high-level information contained in natural language is domain invariant to different tracking domains. Based on this observation, we propose to introduce natural language representation into visual MOT models for boosting the domain generalization ability. However, it is infeasible to label every tracking target with a textual description. To tackle this problem, we design two modules, namely visual context prompting (VCP) and visual-language mixing (VLM). Specifically, VCP generates visual prompts based on the input frames. VLM joints the information in the generated visual prompts and the textual prompts from a pre-defined Trackbook to obtain instance-level pseudo textual description, which is domain invariant to different tracking scenes. Through training models on MOT17 and validating them on MOT20, we observe that the pseudo textual descriptions generated by our proposed modules improve the generalization performance of query-based trackers by large margins.

Jinrong Yang, En Yu, Zeming Li et al.

3D Multi-Object Tracking (MOT) has achieved tremendous achievement thanks to the rapid development of 3D object detection and 2D MOT. Recent advanced works generally employ a series of object attributes, e.g., position, size, velocity, and appearance, to provide the clues for the association in 3D MOT. However, these cues may not be reliable due to some visual noise, such as occlusion and blur, leading to tracking performance bottleneck. To reveal the dilemma, we conduct extensive empirical analysis to expose the key bottleneck of each clue and how they correlate with each other. The analysis results motivate us to efficiently absorb the merits among all cues, and adaptively produce an optimal tacking manner. Specifically, we present Location and Velocity Quality Learning, which efficiently guides the network to estimate the quality of predicted object attributes. Based on these quality estimations, we propose a quality-aware object association (QOA) strategy to leverage the quality score as an important reference factor for achieving robust association. Despite its simplicity, extensive experiments indicate that the proposed strategy significantly boosts tracking performance by 2.2% AMOTA and our method outperforms all existing state-of-the-art works on nuScenes by a large margin. Moreover, QTrack achieves 48.0% and 51.1% AMOTA tracking performance on the nuScenes validation and test sets, which significantly reduces the performance gap between pure camera and LiDAR based trackers.

Zihang Li, Rui Zhou, Yingcheng Shi et al.

When a large language model under reinforcement learning commits a wrong reasoning step early in a trajectory, standard algorithms force it to keep generating until the maximum horizon, spending compute on tokens that never receive positive reward and polluting advantage estimates with post-failure noise. We propose ESPO (Early-Stopping Proximal Policy Optimization), which detects trajectory failure on-the-fly and terminates rollouts early. At each generation step, ESPO computes a surrogate regret using only the logits already computed during sampling, and terminates when the smoothed cumulative regret significantly exceeds its estimated values. Truncated trajectories are treated as absorbing failure states with a terminal reward, concentrating negative temporal-difference (TD) errors near the detected failure step without any additional reward model or human annotation. On DeepSeek-R1-Distill-Qwen-7B trained for mathematical reasoning, ESPO surpasses PPO on AIME~2024 (46.28% vs. 45.25%), AMC~2023 (85.83% vs. 82.94%), and MATH-500 (87.42% vs. 85.43%), while saving more than 20% rollout tokens cumulatively.

Jinrong Yang, Songtao Liu, Zeming Li et al.

The perceptive models of autonomous driving require fast inference within a low latency for safety. While existing works ignore the inevitable environmental changes after processing, streaming perception jointly evaluates the latency and accuracy into a single metric for video online perception, guiding the previous works to search trade-offs between accuracy and speed. In this paper, we explore the performance of real time models on this metric and endow the models with the capacity of predicting the future, significantly improving the results for streaming perception. Specifically, we build a simple framework with two effective modules. One is a Dual Flow Perception module (DFP). It consists of dynamic flow and static flow in parallel to capture moving tendency and basic detection feature, respectively. Trend Aware Loss (TAL) is the other module which adaptively generates loss weight for each object with its moving speed. Realistically, we consider multiple velocities driving scene and further propose Velocity-awared streaming AP (VsAP) to jointly evaluate the accuracy. In this realistic setting, we design a efficient mix-velocity training strategy to guide detector perceive any velocities. Our simple method achieves the state-of-the-art performance on Argoverse-HD dataset and improves the sAP and VsAP by 4.7% and 8.2% respectively compared to the strong baseline, validating its effectiveness.

Hongyu Zhou, Zheng Ge, Weixin Mao et al.

Currently, detecting 3D objects in Bird's-Eye-View (BEV) is superior to other 3D detectors for autonomous driving and robotics. However, transforming image features into BEV necessitates special operators to conduct feature sampling. These operators are not supported on many edge devices, bringing extra obstacles when deploying detectors. To address this problem, we revisit the generation of BEV representation and propose detecting objects in perspective BEV -- a new BEV representation that does not require feature sampling. We demonstrate that perspective BEV features can likewise enjoy the benefits of the BEV paradigm. Moreover, the perspective BEV improves detection performance by addressing issues caused by feature sampling. We propose PersDet for high-performance object detection in perspective BEV space based on this discovery. While implementing a simple and memory-efficient structure, PersDet outperforms existing state-of-the-art monocular methods on the nuScenes benchmark, reaching 34.6% mAP and 40.8% NDS when using ResNet-50 as the backbone.

Weixin Mao, Jinrong Yang, Zheng Ge et al.

Depth perception is a crucial component of monoc-ular 3D detection tasks that typically involve ill-posed problems. In light of the success of sample mining techniques in 2D object detection, we propose a simple yet effective mining strategy for improving depth perception in 3D object detection. Concretely, we introduce a plain metric to evaluate the quality of depth predictions, which chooses the mined sample for the model. Moreover, we propose a Gradient-aware and Model-perceive Mining strategy (GMM) for depth learning, which exploits the predicted depth quality for better depth learning through easy mining. GMM is a general strategy that can be readily applied to several state-of-the-art monocular 3D detectors, improving the accuracy of depth prediction. Extensive experiments on the nuScenes dataset demonstrate that the proposed methods significantly improve the performance of 3D object detection while outperforming other state-of-the-art sample mining techniques by a considerable margin. On the nuScenes benchmark, GMM achieved the state-of-the-art (42.1% mAP and 47.3% NDS) performance in monocular object detection.

Shaoru Wang, Zeming Li, Jin Gao et al.

Self-supervised learning (SSL) has achieved promising downstream performance. However, when facing various resource budgets in real-world applications, it costs a huge computation burden to pretrain multiple networks of various sizes one by one. In this paper, we propose Discriminative-SSL-based Slimmable Pretrained Networks (DSPNet), which can be trained at once and then slimmed to multiple sub-networks of various sizes, each of which faithfully learns good representation and can serve as good initialization for downstream tasks with various resource budgets. Specifically, we extend the idea of slimmable networks to a discriminative SSL paradigm, by integrating SSL and knowledge distillation gracefully. We show comparable or improved performance of DSPNet on ImageNet to the networks individually pretrained one by one under the linear evaluation and semi-supervised evaluation protocols, while reducing large training cost. The pretrained models also generalize well on downstream detection and segmentation tasks. Code will be made public.

Peng Dai, Yang Zhang, Tao Liu et al.

It is especially challenging to achieve real-time human motion tracking on a standalone VR Head-Mounted Display (HMD) such as Meta Quest and PICO. In this paper, we propose HMD-Poser, the first unified approach to recover full-body motions using scalable sparse observations from HMD and body-worn IMUs. In particular, it can support a variety of input scenarios, such as HMD, HMD+2IMUs, HMD+3IMUs, etc. The scalability of inputs may accommodate users' choices for both high tracking accuracy and easy-to-wear. A lightweight temporal-spatial feature learning network is proposed in HMD-Poser to guarantee that the model runs in real-time on HMDs. Furthermore, HMD-Poser presents online body shape estimation to improve the position accuracy of body joints. Extensive experimental results on the challenging AMASS dataset show that HMD-Poser achieves new state-of-the-art results in both accuracy and real-time performance. We also build a new free-dancing motion dataset to evaluate HMD-Poser's on-device performance and investigate the performance gap between synthetic data and real-captured sensor data. Finally, we demonstrate our HMD-Poser with a real-time Avatar-driving application on a commercial HMD. Our code and free-dancing motion dataset are available https://pico-ai-team.github.io/hmd-poser

Huiqun Wang, Yiping Bao, Panwang Pan et al.

Recent advancements in multi-modal pre-training for 3D point clouds have demonstrated promising results by aligning heterogeneous features across 3D shapes and their corresponding 2D images and language descriptions. However, current straightforward solutions often overlook intricate structural relations among samples, potentially limiting the full capabilities of multi-modal learning. To address this issue, we introduce Multi-modal Relation Distillation (MRD), a tri-modal pre-training framework, which is designed to effectively distill reputable large Vision-Language Models (VLM) into 3D backbones. MRD aims to capture both intra-relations within each modality as well as cross-relations between different modalities and produce more discriminative 3D shape representations. Notably, MRD achieves significant improvements in downstream zero-shot classification tasks and cross-modality retrieval tasks, delivering new state-of-the-art performance.

Weiyu Li, Xuanyang Zhang, Zheng Sun et al.

While generative artificial intelligence has advanced significantly across text, image, audio, and video domains, 3D generation remains comparatively underdeveloped due to fundamental challenges such as data scarcity, algorithmic limitations, and ecosystem fragmentation. To this end, we present Step1X-3D, an open framework addressing these challenges through: (1) a rigorous data curation pipeline processing >5M assets to create a 2M high-quality dataset with standardized geometric and textural properties; (2) a two-stage 3D-native architecture combining a hybrid VAE-DiT geometry generator with an diffusion-based texture synthesis module; and (3) the full open-source release of models, training code, and adaptation modules. For geometry generation, the hybrid VAE-DiT component produces TSDF representations by employing perceiver-based latent encoding with sharp edge sampling for detail preservation. The diffusion-based texture synthesis module then ensures cross-view consistency through geometric conditioning and latent-space synchronization. Benchmark results demonstrate state-of-the-art performance that exceeds existing open-source methods, while also achieving competitive quality with proprietary solutions. Notably, the framework uniquely bridges the 2D and 3D generation paradigms by supporting direct transfer of 2D control techniques~(e.g., LoRA) to 3D synthesis. By simultaneously advancing data quality, algorithmic fidelity, and reproducibility, Step1X-3D aims to establish new standards for open research in controllable 3D asset generation.

Mingkai Jia, Mingxiao Li, Liaoyuan Fan et al.

Recent advances in visual generation have highlighted the rise of Latent Generative Models (LGMs), which rely on effective visual tokenizers to bridge pixels and semantics. However, existing tokenizers are typically trained from scratch and struggle to balance semantic representation and reconstruction fidelity, particularly in high-dimensional latent spaces. In this work, we introduce DINO-Tok, a DINO-based visual tokenizer that unifies hierarchical representations into an information-complete latent space. By integrating shallow features that retain fine-grained details with deep features encoding global semantics, DINO-Tok effectively bridges pretrained representations and visual generation. We further analyze the challenges of vector quantization (VQ) in this high-dimensional space, where key information is often lost and codebook collapse occurs. We thus propose a global PCA reweighting mechanism to stabilize VQ and preserve essential information across dimensions. On ImageNet 256$\times$256, DINO-Tok achieves state-of-the-art reconstruction performance, reaching 28.54 PSNR for autoencoding and 23.98 PSNR for VQ-based modeling, significantly outperforming prior tokenizers and comparable to billion-level data trained models (such as Hunyuan and Wan). These results demonstrate that adapting powerful pretrained vision models like DINO for tokenization enables semantically aligned and high-fidelity latent representations, enabling next-generation visual generative models. Code will be publicly available at https://github.com/MKJia/DINO-Tok.

Zeming Li, Xiangyue Liu, Xiangyu Zhang et al.

Diffusion models have emerged as powerful generative frameworks, creating data samples by progressively denoising an initial random state. Traditionally, this initial state is sampled from a simple, fixed distribution like isotropic Gaussian, inherently lacking structure and a direct mechanism for external control. While recent efforts have explored ways to introduce controllability into the diffusion process, particularly at the initialization stage, they often rely on deterministic or heuristic approaches. These methods can be suboptimal, lack expressiveness, and are difficult to scale or integrate into more sophisticated optimization frameworks. In this paper, we introduce NoiseAR, a novel method for AutoRegressive Initial Noise Prior for Diffusion Models. Instead of a static, unstructured source, NoiseAR learns to generate a dynamic and controllable prior distribution for the initial noise. We formulate the generation of the initial noise prior's parameters as an autoregressive probabilistic modeling task over spatial patches or tokens. This approach enables NoiseAR to capture complex spatial dependencies and introduce learned structure into the initial state. Crucially, NoiseAR is designed to be conditional, allowing text prompts to directly influence the learned prior, thereby achieving fine-grained control over the diffusion initialization. Our experiments demonstrate that NoiseAR can generate initial noise priors that lead to improved sample quality and enhanced consistency with conditional inputs, offering a powerful, learned alternative to traditional random initialization. A key advantage of NoiseAR is its probabilistic formulation, which naturally supports seamless integration into probabilistic frameworks like Markov Decision Processes and Reinforcement Learning. Our code will be available at https://github.com/HKUST-SAIL/NoiseAR/

Yanwei Li, Hengshuang Zhao, Xiaojuan Qi et al.

In this paper, we present a conceptually simple, strong, and efficient framework for fully- and weakly-supervised panoptic segmentation, called Panoptic FCN. Our approach aims to represent and predict foreground things and background stuff in a unified fully convolutional pipeline, which can be optimized with point-based fully or weak supervision. In particular, Panoptic FCN encodes each object instance or stuff category with the proposed kernel generator and produces the prediction by convolving the high-resolution feature directly. With this approach, instance-aware and semantically consistent properties for things and stuff can be respectively satisfied in a simple generate-kernel-then-segment workflow. Without extra boxes for localization or instance separation, the proposed approach outperforms the previous box-based and -free models with high efficiency. Furthermore, we propose a new form of point-based annotation for weakly-supervised panoptic segmentation. It only needs several random points for both things and stuff, which dramatically reduces the annotation cost of human. The proposed Panoptic FCN is also proved to have much superior performance in this weakly-supervised setting, which achieves 82% of the fully-supervised performance with only 20 randomly annotated points per instance. Extensive experiments demonstrate the effectiveness and efficiency of Panoptic FCN on COCO, VOC 2012, Cityscapes, and Mapillary Vistas datasets. And it sets up a new leading benchmark for both fully- and weakly-supervised panoptic segmentation. Our code and models are made publicly available at https://github.com/dvlab-research/PanopticFCN.

Songyang Zhang, Lin Song, Songtao Liu et al.

In this report, we introduce our real-time 2D object detection system for the realistic autonomous driving scenario. Our detector is built on a newly designed YOLO model, called YOLOX. On the Argoverse-HD dataset, our system achieves 41.0 streaming AP, which surpassed second place by 7.8/6.1 on detection-only track/fully track, respectively. Moreover, equipped with TensorRT, our model achieves the 30FPS inference speed with a high-resolution input size (e.g., 1440-2304). Code and models will be available at https://github.com/Megvii-BaseDetection/YOLOX

Zheng Ge, Songtao Liu, Feng Wang et al.

In this report, we present some experienced improvements to YOLO series, forming a new high-performance detector -- YOLOX. We switch the YOLO detector to an anchor-free manner and conduct other advanced detection techniques, i.e., a decoupled head and the leading label assignment strategy SimOTA to achieve state-of-the-art results across a large scale range of models: For YOLO-Nano with only 0.91M parameters and 1.08G FLOPs, we get 25.3% AP on COCO, surpassing NanoDet by 1.8% AP; for YOLOv3, one of the most widely used detectors in industry, we boost it to 47.3% AP on COCO, outperforming the current best practice by 3.0% AP; for YOLOX-L with roughly the same amount of parameters as YOLOv4-CSP, YOLOv5-L, we achieve 50.0% AP on COCO at a speed of 68.9 FPS on Tesla V100, exceeding YOLOv5-L by 1.8% AP. Further, we won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021) using a single YOLOX-L model. We hope this report can provide useful experience for developers and researchers in practical scenes, and we also provide deploy versions with ONNX, TensorRT, NCNN, and Openvino supported. Source code is at https://github.com/Megvii-BaseDetection/YOLOX.

Songyang Zhang, Zeming Li, Shipeng Yan et al.

Despite the recent success of deep neural networks, it remains challenging to effectively model the long-tail class distribution in visual recognition tasks. To address this problem, we first investigate the performance bottleneck of the two-stage learning framework via ablative study. Motivated by our discovery, we propose a unified distribution alignment strategy for long-tail visual recognition. Specifically, we develop an adaptive calibration function that enables us to adjust the classification scores for each data point. We then introduce a generalized re-weight method in the two-stage learning to balance the class prior, which provides a flexible and unified solution to diverse scenarios in visual recognition tasks. We validate our method by extensive experiments on four tasks, including image classification, semantic segmentation, object detection, and instance segmentation. Our approach achieves the state-of-the-art results across all four recognition tasks with a simple and unified framework. The code and models will be made publicly available at: https://github.com/Megvii-BaseDetection/DisAlign

Zheng Ge, Songtao Liu, Zeming Li et al.

Recent advances in label assignment in object detection mainly seek to independently define positive/negative training samples for each ground-truth (gt) object. In this paper, we innovatively revisit the label assignment from a global perspective and propose to formulate the assigning procedure as an Optimal Transport (OT) problem -- a well-studied topic in Optimization Theory. Concretely, we define the unit transportation cost between each demander (anchor) and supplier (gt) pair as the weighted summation of their classification and regression losses. After formulation, finding the best assignment solution is converted to solve the optimal transport plan at minimal transportation costs, which can be solved via Sinkhorn-Knopp Iteration. On COCO, a single FCOS-ResNet-50 detector equipped with Optimal Transport Assignment (OTA) can reach 40.7% mAP under 1X scheduler, outperforming all other existing assigning methods. Extensive experiments conducted on COCO and CrowdHuman further validate the effectiveness of our proposed OTA, especially its superiority in crowd scenarios. The code is available at https://github.com/Megvii-BaseDetection/OTA.

Zeming Li, Songtao Liu, Jian Sun

In this paper, we present a novel approach, Momentum$^2$ Teacher, for student-teacher based self-supervised learning. The approach performs momentum update on both network weights and batch normalization (BN) statistics. The teacher's weight is a momentum update of the student, and the teacher's BN statistics is a momentum update of those in history. The Momentum$^2$ Teacher is simple and efficient. It can achieve the state of the art results (74.5\%) under ImageNet linear evaluation protocol using small-batch size(\eg, 128), without requiring large-batch training on special hardware like TPU or inefficient across GPU operation (\eg, shuffling BN, synced BN). Our implementation and pre-trained models will be given on GitHub\footnote{https://github.com/zengarden/momentum2-teacher}.

Jianfeng Wang, Lin Song, Zeming Li et al.

Mainstream object detectors based on the fully convolutional network has achieved impressive performance. While most of them still need a hand-designed non-maximum suppression (NMS) post-processing, which impedes fully end-to-end training. In this paper, we give the analysis of discarding NMS, where the results reveal that a proper label assignment plays a crucial role. To this end, for fully convolutional detectors, we introduce a Prediction-aware One-To-One (POTO) label assignment for classification to enable end-to-end detection, which obtains comparable performance with NMS. Besides, a simple 3D Max Filtering (3DMF) is proposed to utilize the multi-scale features and improve the discriminability of convolutions in the local region. With these techniques, our end-to-end framework achieves competitive performance against many state-of-the-art detectors with NMS on COCO and CrowdHuman datasets. The code is available at https://github.com/Megvii-BaseDetection/DeFCN .

Lin Song, Yanwei Li, Zhengkai Jiang et al.

The Feature Pyramid Network (FPN) presents a remarkable approach to alleviate the scale variance in object representation by performing instance-level assignments. Nevertheless, this strategy ignores the distinct characteristics of different sub-regions in an instance. To this end, we propose a fine-grained dynamic head to conditionally select a pixel-level combination of FPN features from different scales for each instance, which further releases the ability of multi-scale feature representation. Moreover, we design a spatial gate with the new activation function to reduce computational complexity dramatically through spatially sparse convolutions. Extensive experiments demonstrate the effectiveness and efficiency of the proposed method on several state-of-the-art detection benchmarks. Code is available at https://github.com/StevenGrove/DynamicHead.

Lin Song, Yanwei Li, Zhengkai Jiang et al.

The Learnable Tree Filter presents a remarkable approach to model structure-preserving relations for semantic segmentation. Nevertheless, the intrinsic geometric constraint forces it to focus on the regions with close spatial distance, hindering the effective long-range interactions. To relax the geometric constraint, we give the analysis by reformulating it as a Markov Random Field and introduce a learnable unary term. Besides, we propose a learnable spanning tree algorithm to replace the original non-differentiable one, which further improves the flexibility and robustness. With the above improvements, our method can better capture long-range dependencies and preserve structural details with linear complexity, which is extended to several vision tasks for more generic feature transform. Extensive experiments on object detection/instance segmentation demonstrate the consistent improvements over the original version. For semantic segmentation, we achieve leading performance (82.1% mIoU) on the Cityscapes benchmark without bells-and-whistles. Code is available at https://github.com/StevenGrove/LearnableTreeFilterV2.

Yanwei Li, Hengshuang Zhao, Xiaojuan Qi et al.

In this paper, we present a conceptually simple, strong, and efficient framework for panoptic segmentation, called Panoptic FCN. Our approach aims to represent and predict foreground things and background stuff in a unified fully convolutional pipeline. In particular, Panoptic FCN encodes each object instance or stuff category into a specific kernel weight with the proposed kernel generator and produces the prediction by convolving the high-resolution feature directly. With this approach, instance-aware and semantically consistent properties for things and stuff can be respectively satisfied in a simple generate-kernel-then-segment workflow. Without extra boxes for localization or instance separation, the proposed approach outperforms previous box-based and -free models with high efficiency on COCO, Cityscapes, and Mapillary Vistas datasets with single scale input. Our code is made publicly available at https://github.com/Jia-Research-Lab/PanopticFCN.

Han Qiu, Yuchen Ma, Zeming Li et al.

Dense object detectors rely on the sliding-window paradigm that predicts the object over a regular grid of image. Meanwhile, the feature maps on the point of the grid are adopted to generate the bounding box predictions. The point feature is convenient to use but may lack the explicit border information for accurate localization. In this paper, We propose a simple and efficient operator called Border-Align to extract "border features" from the extreme point of the border to enhance the point feature. Based on the BorderAlign, we design a novel detection architecture called BorderDet, which explicitly exploits the border information for stronger classification and more accurate localization. With ResNet-50 backbone, our method improves single-stage detector FCOS by 2.8 AP gains (38.6 v.s. 41.4). With the ResNeXt-101-DCN backbone, our BorderDet obtains 50.3 AP, outperforming the existing state-of-the-art approaches. The code is available at (https://github.com/Megvii-BaseDetection/BorderDet).

Yukang Chen, Peizhen Zhang, Zeming Li et al.

We propose a Dynamic Scale Training paradigm (abbreviated as DST) to mitigate scale variation challenge in object detection. Previous strategies like image pyramid, multi-scale training, and their variants are aiming at preparing scale-invariant data for model optimization. However, the preparation procedure is unaware of the following optimization process that restricts their capability in handling the scale variation. Instead, in our paradigm, we use feedback information from the optimization process to dynamically guide the data preparation. The proposed method is surprisingly simple yet obtains significant gains (2%+ Average Precision on MS COCO dataset), outperforming previous methods. Experimental results demonstrate the efficacy of our proposed DST method towards scale variation handling. It could also generalize to various backbones, benchmarks, and other challenging downstream tasks like instance segmentation. It does not introduce inference overhead and could serve as a free lunch for general detection configurations. Besides, it also facilitates efficient training due to fast convergence. Code and models are available at github.com/yukang2017/Stitcher.

Yanwei Li, Lin Song, Yukang Chen et al.

Recently, numerous handcrafted and searched networks have been applied for semantic segmentation. However, previous works intend to handle inputs with various scales in pre-defined static architectures, such as FCN, U-Net, and DeepLab series. This paper studies a conceptually new method to alleviate the scale variance in semantic representation, named dynamic routing. The proposed framework generates data-dependent routes, adapting to the scale distribution of each image. To this end, a differentiable gating function, called soft conditional gate, is proposed to select scale transform paths on the fly. In addition, the computational cost can be further reduced in an end-to-end manner by giving budget constraints to the gating function. We further relax the network level routing space to support multi-path propagations and skip-connections in each forward, bringing substantial network capacity. To demonstrate the superiority of the dynamic property, we compare with several static architectures, which can be modeled as special cases in the routing space. Extensive experiments are conducted on Cityscapes and PASCAL VOC 2012 to illustrate the effectiveness of the dynamic framework. Code is available at https://github.com/yanwei-li/DynamicRouting.

Lin Song, Yanwei Li, Zeming Li et al.

Learning discriminative global features plays a vital role in semantic segmentation. And most of the existing methods adopt stacks of local convolutions or non-local blocks to capture long-range context. However, due to the absence of spatial structure preservation, these operators ignore the object details when enlarging receptive fields. In this paper, we propose the learnable tree filter to form a generic tree filtering module that leverages the structural property of minimal spanning tree to model long-range dependencies while preserving the details. Furthermore, we propose a highly efficient linear-time algorithm to reduce resource consumption. Thus, the designed modules can be plugged into existing deep neural networks conveniently. To this end, tree filtering modules are embedded to formulate a unified framework for semantic segmentation. We conduct extensive ablation studies to elaborate on the effectiveness and efficiency of the proposed method. Specifically, it attains better performance with much less overhead compared with the classic PSP block and Non-local operation under the same backbone. Our approach is proved to achieve consistent improvements on several benchmarks without bells-and-whistles. Code and models are available at https://github.com/StevenGrove/TreeFilter-Torch.

Zheng Qin, Zeming Li, Zhaoning Zhang et al.

Real-time generic object detection on mobile platforms is a crucial but challenging computer vision task. However, previous CNN-based detectors suffer from enormous computational cost, which hinders them from real-time inference in computation-constrained scenarios. In this paper, we investigate the effectiveness of two-stage detectors in real-time generic detection and propose a lightweight two-stage detector named ThunderNet. In the backbone part, we analyze the drawbacks in previous lightweight backbones and present a lightweight backbone designed for object detection. In the detection part, we exploit an extremely efficient RPN and detection head design. To generate more discriminative feature representation, we design two efficient architecture blocks, Context Enhancement Module and Spatial Attention Module. At last, we investigate the balance between the input resolution, the backbone, and the detection head. Compared with lightweight one-stage detectors, ThunderNet achieves superior performance with only 40% of the computational cost on PASCAL VOC and COCO benchmarks. Without bells and whistles, our model runs at 24.1 fps on an ARM-based device. To the best of our knowledge, this is the first real-time detector reported on ARM platforms. Our code and models are available at \url{https://github.com/qinzheng93/ThunderNet}.

Chenguo Lin, Panwang Pan, Bangbang Yang et al.

Recent advancements in 3D content generation from text or a single image struggle with limited high-quality 3D datasets and inconsistency from 2D multi-view generation. We introduce DiffSplat, a novel 3D generative framework that natively generates 3D Gaussian splats by taming large-scale text-to-image diffusion models. It differs from previous 3D generative models by effectively utilizing web-scale 2D priors while maintaining 3D consistency in a unified model. To bootstrap the training, a lightweight reconstruction model is proposed to instantly produce multi-view Gaussian splat grids for scalable dataset curation. In conjunction with the regular diffusion loss on these grids, a 3D rendering loss is introduced to facilitate 3D coherence across arbitrary views. The compatibility with image diffusion models enables seamless adaptions of numerous techniques for image generation to the 3D realm. Extensive experiments reveal the superiority of DiffSplat in text- and image-conditioned generation tasks and downstream applications. Thorough ablation studies validate the efficacy of each critical design choice and provide insights into the underlying mechanism.

Jin Gao, Shubo Lin, Shaoru Wang et al.

Masked image modeling (MIM) pre-training for large-scale vision transformers (ViTs) has enabled promising downstream performance on top of the learned self-supervised ViT features. In this paper, we question if the \textit{extremely simple} lightweight ViTs' fine-tuning performance can also benefit from this pre-training paradigm, which is considerably less studied yet in contrast to the well-established lightweight architecture design methodology. We use an observation-analysis-solution flow for our study. We first systematically observe different behaviors among the evaluated pre-training methods with respect to the downstream fine-tuning data scales. Furthermore, we analyze the layer representation similarities and attention maps across the obtained models, which clearly show the inferior learning of MIM pre-training on higher layers, leading to unsatisfactory transfer performance on data-insufficient downstream tasks. This finding is naturally a guide to designing our distillation strategies during pre-training to solve the above deterioration problem. Extensive experiments have demonstrated the effectiveness of our approach. Our pre-training with distillation on pure lightweight ViTs with vanilla/hierarchical design ($5.7M$/$6.5M$) can achieve $79.4\%$/$78.9\%$ top-1 accuracy on ImageNet-1K. It also enables SOTA performance on the ADE20K segmentation task ($42.8\%$ mIoU) and LaSOT tracking task ($66.1\%$ AUC) in the lightweight regime. The latter even surpasses all the current SOTA lightweight CPU-realtime trackers.

Taewhan Kim, Hojin Bae, Zeming Li et al.

Visual actionable affordance has emerged as a transformative approach in robotics, focusing on perceiving interaction areas prior to manipulation. Traditional methods rely on pixel sampling to identify successful interaction samples or processing pointclouds for affordance mapping. However, these approaches are computationally intensive and struggle to adapt to diverse and dynamic environments. This paper introduces ManipGPT, a framework designed to predict optimal interaction areas for articulated objects using a large pre-trained vision transformer (ViT). We create a dataset of 9.9k simulated and real images to bridge the visual sim-to-real gap and enhance real-world applicability. By fine-tuning the vision transformer on this small dataset, we significantly improve part-level affordance segmentation, adapting the model's in-context segmentation capabilities to robot manipulation scenarios. This enables effective manipulation across simulated and real-world environments by generating part-level affordance masks, paired with an impedance adaptation policy, sufficiently eliminating the need for complex datasets or perception systems.

Renjie Li, Panwang Pan, Bangbang Yang et al.

The blooming of virtual reality and augmented reality (VR/AR) technologies has driven an increasing demand for the creation of high-quality, immersive, and dynamic environments. However, existing generative techniques either focus solely on dynamic objects or perform outpainting from a single perspective image, failing to meet the requirements of VR/AR applications that need free-viewpoint, 360$^{\circ}$ virtual views where users can move in all directions. In this work, we tackle the challenging task of elevating a single panorama to an immersive 4D experience. For the first time, we demonstrate the capability to generate omnidirectional dynamic scenes with 360$^{\circ}$ views at 4K (4096 $\times$ 2048) resolution, thereby providing an immersive user experience. Our method introduces a pipeline that facilitates natural scene animations and optimizes a set of dynamic Gaussians using efficient splatting techniques for real-time exploration. To overcome the lack of scene-scale annotated 4D data and models, especially in panoramic formats, we propose a novel \textbf{Panoramic Denoiser} that adapts generic 2D diffusion priors to animate consistently in 360$^{\circ}$ images, transforming them into panoramic videos with dynamic scenes at targeted regions. Subsequently, we propose \textbf{Dynamic Panoramic Lifting} to elevate the panoramic video into a 4D immersive environment while preserving spatial and temporal consistency. By transferring prior knowledge from 2D models in the perspective domain to the panoramic domain and the 4D lifting with spatial appearance and geometry regularization, we achieve high-quality Panorama-to-4D generation at a resolution of 4K for the first time.

Panwang Pan, Zhuo Su, Chenguo Lin et al.

Despite recent advancements in high-fidelity human reconstruction techniques, the requirements for densely captured images or time-consuming per-instance optimization significantly hinder their applications in broader scenarios. To tackle these issues, we present HumanSplat which predicts the 3D Gaussian Splatting properties of any human from a single input image in a generalizable manner. In particular, HumanSplat comprises a 2D multi-view diffusion model and a latent reconstruction transformer with human structure priors that adeptly integrate geometric priors and semantic features within a unified framework. A hierarchical loss that incorporates human semantic information is further designed to achieve high-fidelity texture modeling and better constrain the estimated multiple views. Comprehensive experiments on standard benchmarks and in-the-wild images demonstrate that HumanSplat surpasses existing state-of-the-art methods in achieving photorealistic novel-view synthesis.

Zhibo Fan, Yuchen Ma, Zeming Li et al.

Recently few-shot object detection is widely adopted to deal with data-limited situations. While most previous works merely focus on the performance on few-shot categories, we claim that detecting all classes is crucial as test samples may contain any instances in realistic applications, which requires the few-shot detector to learn new concepts without forgetting. Through analysis on transfer learning based methods, some neglected but beneficial properties are utilized to design a simple yet effective few-shot detector, Retentive R-CNN. It consists of Bias-Balanced RPN to debias the pretrained RPN and Re-detector to find few-shot class objects without forgetting previous knowledge. Extensive experiments on few-shot detection benchmarks show that Retentive R-CNN significantly outperforms state-of-the-art methods on overall performance among all settings as it can achieve competitive results on few-shot classes and does not degrade the base class performance at all. Our approach has demonstrated that the long desired never-forgetting learner is available in object detection.

Yuchen Ma, Songtao Liu, Zeming Li et al.

We propose a dense object detector with an instance-wise sampling strategy, named IQDet. Instead of using human prior sampling strategies, we first extract the regional feature of each ground-truth to estimate the instance-wise quality distribution. According to a mixture model in spatial dimensions, the distribution is more noise-robust and adapted to the semantic pattern of each instance. Based on the distribution, we propose a quality sampling strategy, which automatically selects training samples in a probabilistic manner and trains with more high-quality samples. Extensive experiments on MS COCO show that our method steadily improves baseline by nearly 2.4 AP without bells and whistles. Moreover, our best model achieves 51.6 AP, outperforming all existing state-of-the-art one-stage detectors and it is completely cost-free in inference time.

Songtao Liu, Zeming Li, Jian Sun

In this paper, we propose a novel self-supervised representation learning method, Self-EMD, for object detection. Our method directly trained on unlabeled non-iconic image dataset like COCO, instead of commonly used iconic-object image dataset like ImageNet. We keep the convolutional feature maps as the image embedding to preserve spatial structures and adopt Earth Mover's Distance (EMD) to compute the similarity between two embeddings. Our Faster R-CNN (ResNet50-FPN) baseline achieves 39.8% mAP on COCO, which is on par with the state of the art self-supervised methods pre-trained on ImageNet. More importantly, it can be further improved to 40.4% mAP with more unlabeled images, showing its great potential for leveraging more easily obtained unlabeled data. Code will be made available.

Zeming Li, Yuchen Ma, Yukang Chen et al.

In this report, we present our object detection/instance segmentation system, MegDetV2, which works in a two-pass fashion, first to detect instances then to obtain segmentation. Our baseline detector is mainly built on a new designed RPN, called RPN++. On the COCO-2019 detection/instance-segmentation test-dev dataset, our system achieves 61.0/53.1 mAP, which surpassed our 2018 winning results by 5.0/4.2 respectively. We achieve the best results in COCO Challenge 2019 and 2020.

Benjin Zhu, Junqiang Huang, Zeming Li et al.

In this paper, we propose EqCo (Equivalent Rules for Contrastive Learning) to make self-supervised learning irrelevant to the number of negative samples in the contrastive learning framework. Inspired by the InfoMax principle, we point that the margin term in contrastive loss needs to be adaptively scaled according to the number of negative pairs in order to keep steady mutual information bound and gradient magnitude. EqCo bridges the performance gap among a wide range of negative sample sizes, so that for the first time, we can use only a few negative pairs (e.g., 16 per query) to perform self-supervised contrastive training on large-scale vision datasets like ImageNet, while with almost no accuracy drop. This is quite a contrast to the widely used large batch training or memory bank mechanism in current practices. Equipped with EqCo, our simplified MoCo (SiMo) achieves comparable accuracy with MoCov2 on ImageNet (linear evaluation protocol) while only involves 16 negative pairs per query instead of 65536, suggesting that large quantities of negative samples is not a critical factor in contrastive learning frameworks.

Benjin Zhu, Jianfeng Wang, Zhengkai Jiang et al.

Determining positive/negative samples for object detection is known as label assignment. Here we present an anchor-free detector named AutoAssign. It requires little human knowledge and achieves appearance-aware through a fully differentiable weighting mechanism. During training, to both satisfy the prior distribution of data and adapt to category characteristics, we present Center Weighting to adjust the category-specific prior distributions. To adapt to object appearances, Confidence Weighting is proposed to adjust the specific assign strategy of each instance. The two weighting modules are then combined to generate positive and negative weights to adjust each location's confidence. Extensive experiments on the MS COCO show that our method steadily surpasses other best sampling strategies by large margins with various backbones. Moreover, our best model achieves 52.1% AP, outperforming all existing one-stage detectors. Besides, experiments on other datasets, e.g., PASCAL VOC, Objects365, and WiderFace, demonstrate the broad applicability of AutoAssign.