Nanning Zheng

Yichao Shen, Zigang Geng, Yuhui Yuan et al.

We introduce a highly performant 3D object detector for point clouds using the DETR framework. The prior attempts all end up with suboptimal results because they fail to learn accurate inductive biases from the limited scale of training data. In particular, the queries often attend to points that are far away from the target objects, violating the locality principle in object detection. To address the limitation, we introduce a novel 3D Vertex Relative Position Encoding (3DV-RPE) method which computes position encoding for each point based on its relative position to the 3D boxes predicted by the queries in each decoder layer, thus providing clear information to guide the model to focus on points near the objects, in accordance with the principle of locality. In addition, we systematically improve the pipeline from various aspects such as data normalization based on our understanding of the task. We show exceptional results on the challenging ScanNetV2 benchmark, achieving significant improvements over the previous 3DETR in $\rm{AP}_{25}$/$\rm{AP}_{50}$ from 65.0\%/47.0\% to 77.8\%/66.0\%, respectively. In addition, our method sets a new record on ScanNetV2 and SUN RGB-D datasets.Code will be released at http://github.com/yichaoshen-MS/V-DETR.

Shengnan An, Zexiong Ma, Zeqi Lin et al.

Large language models (LLMs) recently exhibited remarkable reasoning capabilities on solving math problems. To further improve their reasoning capabilities, this work explores whether LLMs can LEarn from MistAkes (LEMA), akin to the human learning process. Consider a human student who failed to solve a math problem, he will learn from what mistake he has made and how to correct it. Mimicking this error-driven learning process, LEMA incorporates mistake-correction data pairs during fine-tuning LLMs. Specifically, we first collect inaccurate reasoning paths from various LLMs, and then employ GPT-4 as a ''corrector'' to identify the mistake step, explain the reason for the mistake, correct the mistake and generate the final answer. In addition, we apply a correction-centric evolution strategy that effectively expands the question set for generating correction data. Experiments across various LLMs and reasoning tasks show that LEMA effectively improves CoT-alone fine-tuning. Our further ablations shed light on the non-homogeneous effectiveness between CoT data and correction data. These results suggest a significant potential for LLMs to improve through learning from their mistakes. Our code, models and prompts are publicly available at https://github.com/microsoft/LEMA.

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.

Jiawei Fu, Yanqing Shen, Zhiqiang Jian et al.

Planning and prediction are two important modules of autonomous driving and have experienced tremendous advancement recently. Nevertheless, most existing methods regard planning and prediction as independent and ignore the correlation between them, leading to the lack of consideration for interaction and dynamic changes of traffic scenarios. To address this challenge, we propose InteractionNet, which leverages transformer to share global contextual reasoning among all traffic participants to capture interaction and interconnect planning and prediction to achieve joint. Besides, InteractionNet deploys another transformer to help the model pay extra attention to the perceived region containing critical or unseen vehicles. InteractionNet outperforms other baselines in several benchmarks, especially in terms of safety, which benefits from the joint consideration of planning and forecasting. The code will be available at https://github.com/fujiawei0724/InteractionNet.

Long Chen, Yuchen Li, Chao Huang et al.

Interest in autonomous driving (AD) and intelligent vehicles (IVs) is growing at a rapid pace due to the convenience, safety, and economic benefits. Although a number of surveys have reviewed research achievements in this field, they are still limited in specific tasks, lack of systematic summary and research directions in the future. Here we propose a Survey of Surveys (SoS) for total technologies of AD and IVs that reviews the history, summarizes the milestones, and provides the perspectives, ethics, and future research directions. To our knowledge, this article is the first SoS with milestones in AD and IVs, which constitutes our complete research work together with two other technical surveys. We anticipate that this article will bring novel and diverse insights to researchers and abecedarians, and serve as a bridge between past and future.

Shengnan An, Yifei Li, Zeqi Lin et al. · pku

Recently the prompt-tuning paradigm has attracted significant attention. By only tuning continuous prompts with a frozen pre-trained language model (PLM), prompt-tuning takes a step towards deploying a shared frozen PLM to serve numerous downstream tasks. Although prompt-tuning shows good performance on certain natural language understanding (NLU) tasks, its effectiveness on natural language generation (NLG) tasks is still under-explored. In this paper, we argue that one of the factors hindering the development of prompt-tuning on NLG tasks is the unfamiliar inputs (i.e., inputs are linguistically different from the pretraining corpus). For example, our preliminary exploration reveals a large performance gap between prompt-tuning and fine-tuning when unfamiliar inputs occur frequently in NLG tasks. This motivates us to propose input-tuning, which fine-tunes both the continuous prompts and the input representations, leading to a more effective way to adapt unfamiliar inputs to frozen PLMs. Our proposed input-tuning is conceptually simple and empirically powerful. Experimental results on seven NLG tasks demonstrate that input-tuning is significantly and consistently better than prompt-tuning. Furthermore, on three of these tasks, input-tuning can achieve a comparable or even better performance than fine-tuning.

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.

Yutong Lin, Yuhui Yuan, Zheng Zhang et al.

This paper presents an improved DETR detector that maintains a "plain" nature: using a single-scale feature map and global cross-attention calculations without specific locality constraints, in contrast to previous leading DETR-based detectors that reintroduce architectural inductive biases of multi-scale and locality into the decoder. We show that two simple technologies are surprisingly effective within a plain design to compensate for the lack of multi-scale feature maps and locality constraints. The first is a box-to-pixel relative position bias (BoxRPB) term added to the cross-attention formulation, which well guides each query to attend to the corresponding object region while also providing encoding flexibility. The second is masked image modeling (MIM)-based backbone pre-training which helps learn representation with fine-grained localization ability and proves crucial for remedying dependencies on the multi-scale feature maps. By incorporating these technologies and recent advancements in training and problem formation, the improved "plain" DETR showed exceptional improvements over the original DETR detector. By leveraging the Object365 dataset for pre-training, it achieved 63.9 mAP accuracy using a Swin-L backbone, which is highly competitive with state-of-the-art detectors which all heavily rely on multi-scale feature maps and region-based feature extraction. Code is available at https://github.com/impiga/Plain-DETR .

Long Chen, Siyu Teng, Bai Li et al.

Growing interest in autonomous driving (AD) and intelligent vehicles (IVs) is fueled by their promise for enhanced safety, efficiency, and economic benefits. While previous surveys have captured progress in this field, a comprehensive and forward-looking summary is needed. Our work fills this gap through three distinct articles. The first part, a "Survey of Surveys" (SoS), outlines the history, surveys, ethics, and future directions of AD and IV technologies. The second part, "Milestones in Autonomous Driving and Intelligent Vehicles Part I: Control, Computing System Design, Communication, HD Map, Testing, and Human Behaviors" delves into the development of control, computing system, communication, HD map, testing, and human behaviors in IVs. This part, the third part, reviews perception and planning in the context of IVs. Aiming to provide a comprehensive overview of the latest advancements in AD and IVs, this work caters to both newcomers and seasoned researchers. By integrating the SoS and Part I, we offer unique insights and strive to serve as a bridge between past achievements and future possibilities in this dynamic field.

Liushuai Shi, Le Wang, Chengjiang Long et al.

Understanding the multiple socially-acceptable future behaviors is an essential task for many vision applications. In this paper, we propose a tree-based method, termed as Social Interpretable Tree (SIT), to address this multi-modal prediction task, where a hand-crafted tree is built depending on the prior information of observed trajectory to model multiple future trajectories. Specifically, a path in the tree from the root to leaf represents an individual possible future trajectory. SIT employs a coarse-to-fine optimization strategy, in which the tree is first built by high-order velocity to balance the complexity and coverage of the tree and then optimized greedily to encourage multimodality. Finally, a teacher-forcing refining operation is used to predict the final fine trajectory. Compared with prior methods which leverage implicit latent variables to represent possible future trajectories, the path in the tree can explicitly explain the rough moving behaviors (e.g., go straight and then turn right), and thus provides better interpretability. Despite the hand-crafted tree, the experimental results on ETH-UCY and Stanford Drone datasets demonstrate that our method is capable of matching or exceeding the performance of state-of-the-art methods. Interestingly, the experiments show that the raw built tree without training outperforms many prior deep neural network based approaches. Meanwhile, our method presents sufficient flexibility in long-term prediction and different best-of-$K$ predictions.

Yulong Liu, Yongqiang Ma, Wei Zhou et al.

Due to the lack of paired samples and the low signal-to-noise ratio of functional MRI (fMRI) signals, reconstructing perceived natural images or decoding their semantic contents from fMRI data are challenging tasks. In this work, we propose, for the first time, a task-agnostic fMRI-based brain decoding model, BrainCLIP, which leverages CLIP's cross-modal generalization ability to bridge the modality gap between brain activity, image, and text. Our experiments demonstrate that CLIP can act as a pivot for generic brain decoding tasks, including zero-shot visual categories decoding, fMRI-image/text matching, and fMRI-to-image generation. Specifically, BrainCLIP aims to train a mapping network that transforms fMRI patterns into a well-aligned CLIP embedding space by combining visual and textual supervision. Our experiments show that this combination can boost the decoding model's performance on certain tasks like fMRI-text matching and fMRI-to-image generation. On the zero-shot visual category decoding task, BrainCLIP achieves significantly better performance than BraVL, a recently proposed multi-modal method specifically designed for this task. BrainCLIP can also reconstruct visual stimuli with high semantic fidelity and establishes a new state-of-the-art for fMRI-based natural image reconstruction in terms of high-level semantic features.

Chao Xia, Chenfeng Xu, Patrick Rim et al.

Current LiDAR odometry, mapping and localization methods leverage point-wise representations of 3D scenes and achieve high accuracy in autonomous driving tasks. However, the space-inefficiency of methods that use point-wise representations limits their development and usage in practical applications. In particular, scan-submap matching and global map representation methods are restricted by the inefficiency of nearest neighbor searching (NNS) for large-volume point clouds. To improve space-time efficiency, we propose a novel method of describing scenes using quadric surfaces, which are far more compact representations of 3D objects than conventional point clouds. In contrast to point cloud-based methods, our quadric representation-based method decomposes a 3D scene into a collection of sparse quadric patches, which improves storage efficiency and avoids the slow point-wise NNS process. Our method first segments a given point cloud into patches and fits each of them to a quadric implicit function. Each function is then coupled with other geometric descriptors of the patch, such as its center position and covariance matrix. Collectively, these patch representations fully describe a 3D scene, which can be used in place of the original point cloud and employed in LiDAR odometry, mapping and localization algorithms. We further design a novel incremental growing method for quadric representations, which eliminates the need to repeatedly re-fit quadric surfaces from the original point cloud. Extensive odometry, mapping and localization experiments on large-volume point clouds in the KITTI and UrbanLoco datasets demonstrate that our method maintains low latency and memory utility while achieving competitive, and even superior, accuracy.

Shaoning Li, Mingyu Li, Yusong Wang et al.

Investigating conformational landscapes of proteins is a crucial way to understand their biological functions and properties. AlphaFlow stands out as a sequence-conditioned generative model that introduces flexibility into structure prediction models by fine-tuning AlphaFold under the flow-matching framework. Despite the advantages of efficient sampling afforded by flow-matching, AlphaFlow still requires multiple runs of AlphaFold to finally generate one single conformation. Due to the heavy consumption of AlphaFold, its applicability is limited in sampling larger set of protein ensembles or the longer chains within a constrained timeframe. In this work, we propose a feature-conditioned generative model called AlphaFlow-Lit to realize efficient protein ensembles generation. In contrast to the full fine-tuning on the entire structure, we focus solely on the light-weight structure module to reconstruct the conformation. AlphaFlow-Lit performs on-par with AlphaFlow and surpasses its distilled version without pretraining, all while achieving a significant sampling acceleration of around 47 times. The advancement in efficiency showcases the potential of AlphaFlow-Lit in enabling faster and more scalable generation of protein ensembles.

Wenjie Gao, Jiawei Fu, Yanqing Shen et al.

High-definition (HD) maps play a crucial role in autonomous driving systems. Recent methods have attempted to construct HD maps in real-time using vehicle onboard sensors. Due to the inherent limitations of onboard sensors, which include sensitivity to detection range and susceptibility to occlusion by nearby vehicles, the performance of these methods significantly declines in complex scenarios and long-range detection tasks. In this paper, we explore a new perspective that boosts HD map construction through the use of satellite maps to complement onboard sensors. We initially generate the satellite map tiles for each sample in nuScenes and release a complementary dataset for further research. To enable better integration of satellite maps with existing methods, we propose a hierarchical fusion module, which includes feature-level fusion and BEV-level fusion. The feature-level fusion, composed of a mask generator and a masked cross-attention mechanism, is used to refine the features from onboard sensors. The BEV-level fusion mitigates the coordinate differences between features obtained from onboard sensors and satellite maps through an alignment module. The experimental results on the augmented nuScenes showcase the seamless integration of our module into three existing HD map construction methods. The satellite maps and our proposed module notably enhance their performance in both HD map semantic segmentation and instance detection tasks.

Tao Yang, Yuwang Wang, Yan Lv et al.

Targeting to understand the underlying explainable factors behind observations and modeling the conditional generation process on these factors, we connect disentangled representation learning to Diffusion Probabilistic Models (DPMs) to take advantage of the remarkable modeling ability of DPMs. We propose a new task, disentanglement of (DPMs): given a pre-trained DPM, without any annotations of the factors, the task is to automatically discover the inherent factors behind the observations and disentangle the gradient fields of DPM into sub-gradient fields, each conditioned on the representation of each discovered factor. With disentangled DPMs, those inherent factors can be automatically discovered, explicitly represented, and clearly injected into the diffusion process via the sub-gradient fields. To tackle this task, we devise an unsupervised approach named DisDiff, achieving disentangled representation learning in the framework of DPMs. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of DisDiff.

Kun Xia, Le Wang, Sanping Zhou et al.

The main challenge of Temporal Action Localization is to retrieve subtle human actions from various co-occurring ingredients, e.g., context and background, in an untrimmed video. While prior approaches have achieved substantial progress through devising advanced action detectors, they still suffer from these co-occurring ingredients which often dominate the actual action content in videos. In this paper, we explore two orthogonal but complementary aspects of a video snippet, i.e., the action features and the co-occurrence features. Especially, we develop a novel auxiliary task by decoupling these two types of features within a video snippet and recombining them to generate a new feature representation with more salient action information for accurate action localization. We term our method RefactorNet, which first explicitly factorizes the action content and regularizes its co-occurrence features, and then synthesizes a new action-dominated video representation. Extensive experimental results and ablation studies on THUMOS14 and ActivityNet v1.3 demonstrate that our new representation, combined with a simple action detector, can significantly improve the action localization performance.

Chengrui Wei, Meng Yang, Lei He et al.

It has long been an ill-posed problem to predict absolute depth maps from single images in real (unseen) indoor scenes. We observe that it is essentially due to not only the scale-ambiguous problem but also the focal-ambiguous problem that decreases the generalization ability of monocular depth estimation. That is, images may be captured by cameras of different focal lengths in scenes of different scales. In this paper, we develop a focal-and-scale depth estimation model to well learn absolute depth maps from single images in unseen indoor scenes. First, a relative depth estimation network is adopted to learn relative depths from single images with diverse scales/semantics. Second, multi-scale features are generated by mapping a single focal length value to focal length features and concatenating them with intermediate features of different scales in relative depth estimation. Finally, relative depths and multi-scale features are jointly fed into an absolute depth estimation network. In addition, a new pipeline is developed to augment the diversity of focal lengths of public datasets, which are often captured with cameras of the same or similar focal lengths. Our model is trained on augmented NYUDv2 and tested on three unseen datasets. Our model considerably improves the generalization ability of depth estimation by 41%/13% (RMSE) with/without data augmentation compared with five recent SOTAs and well alleviates the deformation problem in 3D reconstruction. Notably, our model well maintains the accuracy of depth estimation on original NYUDv2.

Yutong Lin, Ze Liu, Zheng Zhang et al.

Frozen pretrained models have become a viable alternative to the pretraining-then-finetuning paradigm for transfer learning. However, with frozen models there are relatively few parameters available for adapting to downstream tasks, which is problematic in computer vision where tasks vary significantly in input/output format and the type of information that is of value. In this paper, we present a study of frozen pretrained models when applied to diverse and representative computer vision tasks, including object detection, semantic segmentation and video action recognition. From this empirical analysis, our work answers the questions of what pretraining task fits best with this frozen setting, how to make the frozen setting more flexible to various downstream tasks, and the effect of larger model sizes. We additionally examine the upper bound of performance using a giant frozen pretrained model with 3 billion parameters (SwinV2-G) and find that it reaches competitive performance on a varied set of major benchmarks with only one shared frozen base network: 60.0 box mAP and 52.2 mask mAP on COCO object detection test-dev, 57.6 val mIoU on ADE20K semantic segmentation, and 81.7 top-1 accuracy on Kinetics-400 action recognition. With this work, we hope to bring greater attention to this promising path of freezing pretrained image models.

Xingyu Chen, Jianru Xue, Jianwu Fang et al.

In this paper, we propose a lightweight system, RDS-SLAM, based on ORB-SLAM2, which can accurately estimate poses and build semantic maps at object level for dynamic scenarios in real time using only one commonly used Intel Core i7 CPU. In RDS-SLAM, three major improvements, as well as major architectural modifications, are proposed to overcome the limitations of ORB-SLAM2. Firstly, it adopts a lightweight object detection neural network in key frames. Secondly, an efficient tracking and prediction mechanism is embedded into the system to remove the feature points belonging to movable objects in all incoming frames. Thirdly, a semantic octree map is built by probabilistic fusion of detection and tracking results, which enables a robot to maintain a semantic description at object level for potential interactions in dynamic scenarios. We evaluate RDS-SLAM in TUM RGB-D dataset, and experimental results show that RDS-SLAM can run with 30.3 ms per frame in dynamic scenarios using only an Intel Core i7 CPU, and achieves comparable accuracy compared with the state-of-the-art SLAM systems which heavily rely on both Intel Core i7 CPUs and powerful GPUs.

Ning Gao, Sanping Zhou, Le Wang et al.

Semi-supervised learning has emerged as a widely adopted technique in the field of medical image segmentation. The existing works either focuses on the construction of consistency constraints or the generation of pseudo labels to provide high-quality supervisory signals, whose main challenge mainly comes from how to keep the continuous improvement of model capabilities. In this paper, we propose a simple yet effective semi-supervised learning framework, termed Progressive Mean Teachers (PMT), for medical image segmentation, whose goal is to generate high-fidelity pseudo labels by learning robust and diverse features in the training process. Specifically, our PMT employs a standard mean teacher to penalize the consistency of the current state and utilizes two sets of MT architectures for co-training. The two sets of MT architectures are individually updated for prolonged periods to maintain stable model diversity established through performance gaps generated by iteration differences. Additionally, a difference-driven alignment regularizer is employed to expedite the alignment of lagging models with the representation capabilities of leading models. Furthermore, a simple yet effective pseudo-label filtering algorithm is employed for facile evaluation of models and selection of high-fidelity pseudo-labels outputted when models are operating at high performance for co-training purposes. Experimental results on two datasets with different modalities, i.e., CT and MRI, demonstrate that our method outperforms the state-of-the-art medical image segmentation approaches across various dimensions. The code is available at https://github.com/Axi404/PMT.

Jingwen Fu, Tao Yang, Yuwang Wang et al.

In-context learning, i.e., learning from context examples, is an impressive ability of Transformer. Training Transformers to possess this in-context learning skill is computationally intensive due to the occurrence of learning plateaus, which are periods within the training process where there is minimal or no enhancement in the model's in-context learning capability. To study the mechanism behind the learning plateaus, we conceptually seperate a component within the model's internal representation that is exclusively affected by the model's weights. We call this the "weights component", and the remainder is identified as the "context component". By conducting meticulous and controlled experiments on synthetic tasks, we note that the persistence of learning plateaus correlates with compromised functionality of the weights component. Recognizing the impaired performance of the weights component as a fundamental behavior drives learning plateaus, we have developed three strategies to expedite the learning of Transformers. The effectiveness of these strategies is further confirmed in natural language processing tasks. In conclusion, our research demonstrates the feasibility of cultivating a powerful in-context learning ability within AI systems in an eco-friendly manner.

Zhiqin Yang, Yuhan Liu, Jingwen Fu et al.

Although natural language is the default medium for Large Language Models (LLMs), its limited expressive capacity creates a profound bottleneck for complex problem-solving. While recent advancements in AI have relied heavily on scaling, merely internalizing knowledge does not guarantee its effective application. Defining language representation as the linguistic and symbolic constructs used to map and model the real world, this paper argues that shaping schemas through advanced language representation is the next frontier for expanding LLM intelligence. We posit that an LLM's knowledge activation and organization -- its schema -- depends heavily on the structural and symbolic sophistication of the language used to represent a given task. This paper contributes both a formalization of this claim and the empirical evidence to support it. With a new formalization, we present multiple lines of evidence to support our position: Firstly, we review recent empirical practices and emerging methodologies that demonstrate the substantial performance gains achievable through deliberate language representation design, even without modifying model parameters or scale. Secondly, we conduct controlled experiments showing that LLM performance and its internal feature activations vary under different language representations of the same underlying task. Together, these findings highlight language representation design as a promising direction for future research.

Bohan Wang, Jingwen Fu, Huishuai Zhang et al.

Recently, Arjevani et al. [1] established a lower bound of iteration complexity for the first-order optimization under an $L$-smooth condition and a bounded noise variance assumption. However, a thorough review of existing literature on Adam's convergence reveals a noticeable gap: none of them meet the above lower bound. In this paper, we close the gap by deriving a new convergence guarantee of Adam, with only an $L$-smooth condition and a bounded noise variance assumption. Our results remain valid across a broad spectrum of hyperparameters. Especially with properly chosen hyperparameters, we derive an upper bound of the iteration complexity of Adam and show that it meets the lower bound for first-order optimizers. To the best of our knowledge, this is the first to establish such a tight upper bound for Adam's convergence. Our proof utilizes novel techniques to handle the entanglement between momentum and adaptive learning rate and to convert the first-order term in the Descent Lemma to the gradient norm, which may be of independent interest.

He Zhang, Siyuan Liu, Jiacheng You et al.

Orbital-free density functional theory (OFDFT) is a quantum chemistry formulation that has a lower cost scaling than the prevailing Kohn-Sham DFT, which is increasingly desired for contemporary molecular research. However, its accuracy is limited by the kinetic energy density functional, which is notoriously hard to approximate for non-periodic molecular systems. Here we propose M-OFDFT, an OFDFT approach capable of solving molecular systems using a deep learning functional model. We build the essential non-locality into the model, which is made affordable by the concise density representation as expansion coefficients under an atomic basis. With techniques to address unconventional learning challenges therein, M-OFDFT achieves a comparable accuracy with Kohn-Sham DFT on a wide range of molecules untouched by OFDFT before. More attractively, M-OFDFT extrapolates well to molecules much larger than those seen in training, which unleashes the appealing scaling of OFDFT for studying large molecules including proteins, representing an advancement of the accuracy-efficiency trade-off frontier in quantum chemistry.

Jingwen Fu, Bohan Wang, Huishuai Zhang et al.

Momentum has become a crucial component in deep learning optimizers, necessitating a comprehensive understanding of when and why it accelerates stochastic gradient descent (SGD). To address the question of ''when'', we establish a meaningful comparison framework that examines the performance of SGD with Momentum (SGDM) under the \emph{effective learning rates} $η_{ef}$, a notion unifying the influence of momentum coefficient $μ$ and batch size $b$ over learning rate $η$. In the comparison of SGDM and SGD with the same effective learning rate and the same batch size, we observe a consistent pattern: when $η_{ef}$ is small, SGDM and SGD experience almost the same empirical training losses; when $η_{ef}$ surpasses a certain threshold, SGDM begins to perform better than SGD. Furthermore, we observe that the advantage of SGDM over SGD becomes more pronounced with a larger batch size. For the question of ``why'', we find that the momentum acceleration is closely related to \emph{abrupt sharpening} which is to describe a sudden jump of the directional Hessian along the update direction. Specifically, the misalignment between SGD and SGDM happens at the same moment that SGD experiences abrupt sharpening and converges slower. Momentum improves the performance of SGDM by preventing or deferring the occurrence of abrupt sharpening. Together, this study unveils the interplay between momentum, learning rates, and batch sizes, thus improving our understanding of momentum acceleration.

Yizhe Li, Sanping Zhou, Zheng Qin et al.

Multi-Object Tracking (MOT) remains a vital component of intelligent video analysis, which aims to locate targets and maintain a consistent identity for each target throughout a video sequence. Existing works usually learn a discriminative feature representation, such as motion and appearance, to associate the detections across frames, which are easily affected by mutual occlusion and background clutter in practice. In this paper, we propose a simple yet effective two-stage feature learning paradigm to jointly learn single-shot and multi-shot features for different targets, so as to achieve robust data association in the tracking process. For the detections without being associated, we design a novel single-shot feature learning module to extract discriminative features of each detection, which can efficiently associate targets between adjacent frames. For the tracklets being lost several frames, we design a novel multi-shot feature learning module to extract discriminative features of each tracklet, which can accurately refind these lost targets after a long period. Once equipped with a simple data association logic, the resulting VisualTracker can perform robust MOT based on the single-shot and multi-shot feature representations. Extensive experimental results demonstrate that our method has achieved significant improvements on MOT17 and MOT20 datasets while reaching state-of-the-art performance on DanceTrack dataset.

Hao Zhang, Lumin Xu, Shenqi Lai et al.

Current image-based keypoint detection methods for animal (including human) bodies and faces are generally divided into full-supervised and few-shot class-agnostic approaches. The former typically relies on laborious and time-consuming manual annotations, posing considerable challenges in expanding keypoint detection to a broader range of keypoint categories and animal species. The latter, though less dependent on extensive manual input, still requires necessary support images with annotation for reference during testing. To realize zero-shot keypoint detection without any prior annotation, we introduce the Open-Vocabulary Keypoint Detection (OVKD) task, which is innovatively designed to use text prompts for identifying arbitrary keypoints across any species. In pursuit of this goal, we have developed a novel framework named Open-Vocabulary Keypoint Detection with Semantic-feature Matching (KDSM). This framework synergistically combines vision and language models, creating an interplay between language features and local keypoint visual features. KDSM enhances its capabilities by integrating Domain Distribution Matrix Matching (DDMM) and other special modules, such as the Vision-Keypoint Relational Awareness (VKRA) module, improving the framework's generalizability and overall performance.Our comprehensive experiments demonstrate that KDSM significantly outperforms the baseline in terms of performance and achieves remarkable success in the OVKD task.Impressively, our method, operating in a zero-shot fashion, still yields results comparable to state-of-the-art few-shot species class-agnostic keypoint detection methods.We will make the source code publicly accessible.

Shengnan An, Zeqi Lin, Bei Chen et al.

Abstraction is a desirable capability for deep learning models, which means to induce abstract concepts from concrete instances and flexibly apply them beyond the learning context. At the same time, there is a lack of clear understanding about both the presence and further characteristics of this capability in deep learning models. In this paper, we introduce a systematic probing framework to explore the abstraction capability of deep learning models from a transferability perspective. A set of controlled experiments are conducted based on this framework, providing strong evidence that two probed pre-trained language models (PLMs), T5 and GPT2, have the abstraction capability. We also conduct in-depth analysis, thus shedding further light: (1) the whole training phase exhibits a "memorize-then-abstract" two-stage process; (2) the learned abstract concepts are gathered in a few middle-layer attention heads, rather than being evenly distributed throughout the model; (3) the probed abstraction capabilities exhibit robustness against concept mutations, and are more robust to low-level/source-side mutations than high-level/target-side ones; (4) generic pre-training is critical to the emergence of abstraction capability, and PLMs exhibit better abstraction with larger model sizes and data scales.

Yanqing Shen, Sanping Zhou, Jingwen Fu et al.

Visual place recognition (VPR) is usually considered as a specific image retrieval problem. Limited by existing training frameworks, most deep learning-based works cannot extract sufficiently stable global features from RGB images and rely on a time-consuming re-ranking step to exploit spatial structural information for better performance. In this paper, we propose StructVPR, a novel training architecture for VPR, to enhance structural knowledge in RGB global features and thus improve feature stability in a constantly changing environment. Specifically, StructVPR uses segmentation images as a more definitive source of structural knowledge input into a CNN network and applies knowledge distillation to avoid online segmentation and inference of seg-branch in testing. Considering that not all samples contain high-quality and helpful knowledge, and some even hurt the performance of distillation, we partition samples and weigh each sample's distillation loss to enhance the expected knowledge precisely. Finally, StructVPR achieves impressive performance on several benchmarks using only global retrieval and even outperforms many two-stage approaches by a large margin. After adding additional re-ranking, ours achieves state-of-the-art performance while maintaining a low computational cost.

Jingjing Jiang, Ziyi Liu, Nanning Zheng

Benefiting from large-scale pretrained vision language models (VLMs), the performance of visual question answering (VQA) has approached human oracles. However, finetuning such models on limited data often suffers from overfitting and poor generalization issues, leading to a lack of model robustness. In this paper, we aim to improve input robustness from an information bottleneck perspective when adapting pretrained VLMs to the downstream VQA task. Input robustness refers to the ability of models to defend against visual and linguistic input variations, as well as shortcut learning involved in inputs. Generally, the representations obtained by pretrained VLMs inevitably contain irrelevant and redundant information for a specific downstream task, resulting in statistically spurious correlations and insensitivity to input variations. To encourage representations to converge to a minimal sufficient statistic in multimodal learning, we propose Correlation Information Bottleneck (CIB), which seeks a tradeoff between compression and redundancy in representations by minimizing the mutual information (MI) between inputs and representations while maximizing the MI between outputs and representations. Moreover, we derive a tight theoretical upper bound for the mutual information between multimodal inputs and representations, incorporating different internal correlations that guide models to learn more robust representations and facilitate modality alignment. Extensive experiments consistently demonstrate the effectiveness and superiority of the proposed CIB in terms of input robustness and accuracy.

Tao Yang, Yuwang Wang, Yan Lu et al.

Obtaining the human-like perception ability of abstracting visual concepts from concrete pixels has always been a fundamental and important target in machine learning research fields such as disentangled representation learning and scene decomposition. Towards this goal, we propose an unsupervised transformer-based Visual Concepts Tokenization framework, dubbed VCT, to perceive an image into a set of disentangled visual concept tokens, with each concept token responding to one type of independent visual concept. Particularly, to obtain these concept tokens, we only use cross-attention to extract visual information from the image tokens layer by layer without self-attention between concept tokens, preventing information leakage across concept tokens. We further propose a Concept Disentangling Loss to facilitate that different concept tokens represent independent visual concepts. The cross-attention and disentangling loss play the role of induction and mutual exclusion for the concept tokens, respectively. Extensive experiments on several popular datasets verify the effectiveness of VCT on the tasks of disentangled representation learning and scene decomposition. VCT achieves the state of the art results by a large margin.

Jingwen Fu, Zhizheng Zhang, Dacheng Yin et al.

This paper explores the connection between learning trajectories of Deep Neural Networks (DNNs) and their generalization capabilities when optimized using (stochastic) gradient descent algorithms. Instead of concentrating solely on the generalization error of the DNN post-training, we present a novel perspective for analyzing generalization error by investigating the contribution of each update step to the change in generalization error. This perspective allows for a more direct comprehension of how the learning trajectory influences generalization error. Building upon this analysis, we propose a new generalization bound that incorporates more extensive trajectory information. Our proposed generalization bound depends on the complexity of learning trajectory and the ratio between the bias and diversity of training set. Experimental findings reveal that our method effectively captures the generalization error throughout the training process. Furthermore, our approach can also track changes in generalization error when adjustments are made to learning rates and label noise levels. These results demonstrate that learning trajectory information is a valuable indicator of a model's generalization capabilities.

Tao Yang, Shenglong Zhou, Yuwang Wang et al.

Deep neural networks often suffer the data distribution shift between training and testing, and the batch statistics are observed to reflect the shift. In this paper, targeting of alleviating distribution shift in test time, we revisit the batch normalization (BN) in the training process and reveals two key insights benefiting test-time optimization: $(i)$ preserving the same gradient backpropagation form as training, and $(ii)$ using dataset-level statistics for robust optimization and inference. Based on the two insights, we propose a novel test-time BN layer design, GpreBN, which is optimized during testing by minimizing Entropy loss. We verify the effectiveness of our method on two typical settings with distribution shift, i.e., domain generalization and robustness tasks. Our GpreBN significantly improves the test-time performance and achieves the state of the art results.

Jingjing Jiang, Nanning Zheng

Recently, finetuning pretrained Vision-Language Models (VLMs) has been a prevailing paradigm for achieving state-of-the-art performance in Visual Question Answering (VQA). However, as VLMs scale, finetuning full model parameters for a given task in low-resource settings becomes computationally expensive, storage inefficient, and prone to overfitting. Current parameter-efficient tuning methods dramatically reduce the number of tunable parameters, but there still exists a significant performance gap with full finetuning. In this paper, we propose MixPHM, a redundancy-aware parameter-efficient tuning method that outperforms full finetuning in low-resource VQA. Specifically, MixPHM is a lightweight module implemented by multiple PHM-experts in a mixture-of-experts manner. To reduce parameter redundancy, MixPHM reparameterizes expert weights in a low-rank subspace and shares part of the weights inside and across experts. Moreover, based on a quantitative redundancy analysis for adapters, we propose Redundancy Regularization to reduce task-irrelevant redundancy while promoting task-relevant correlation in MixPHM representations. Experiments conducted on VQA v2, GQA, and OK-VQA demonstrate that MixPHM outperforms state-of-the-art parameter-efficient methods and is the only one consistently surpassing full finetuning.

Zewei Lin, Yanqing Shen, Sanping Zhou et al.

In this paper, we propose a novel and effective Multi-Level Fusion network, named as MLF-DET, for high-performance cross-modal 3D object DETection, which integrates both the feature-level fusion and decision-level fusion to fully utilize the information in the image. For the feature-level fusion, we present the Multi-scale Voxel Image fusion (MVI) module, which densely aligns multi-scale voxel features with image features. For the decision-level fusion, we propose the lightweight Feature-cued Confidence Rectification (FCR) module which further exploits image semantics to rectify the confidence of detection candidates. Besides, we design an effective data augmentation strategy termed Occlusion-aware GT Sampling (OGS) to reserve more sampled objects in the training scenes, so as to reduce overfitting. Extensive experiments on the KITTI dataset demonstrate the effectiveness of our method. Notably, on the extremely competitive KITTI car 3D object detection benchmark, our method reaches 82.89% moderate AP and achieves state-of-the-art performance without bells and whistles.

Yuhao Huang, Sanping Zhou, Junjie Zhang et al.

Efficient representation of point clouds is fundamental for LiDAR-based 3D object detection. While recent grid-based detectors often encode point clouds into either voxels or pillars, the distinctions between these approaches remain underexplored. In this paper, we quantify the differences between the current encoding paradigms and highlight the limited vertical learning within. To tackle these limitations, we introduce a hybrid Voxel-Pillar Fusion network (VPF), which synergistically combines the unique strengths of both voxels and pillars. Specifically, we first develop a sparse voxel-pillar encoder that encodes point clouds into voxel and pillar features through 3D and 2D sparse convolutions respectively, and then introduce the Sparse Fusion Layer (SFL), facilitating bidirectional interaction between sparse voxel and pillar features. Our efficient, fully sparse method can be seamlessly integrated into both dense and sparse detectors. Leveraging this powerful yet straightforward framework, VPF delivers competitive performance, achieving real-time inference speeds on the nuScenes and Waymo Open Dataset. The code will be available.

Yusong Wang, Chaoran Cheng, Shaoning Li et al.

Geometric graph neural networks (GNNs) have emerged as powerful tools for modeling molecular geometry. However, they encounter limitations in effectively capturing long-range interactions in large molecular systems. To address this challenge, we introduce Neural P$^3$M, a versatile enhancer of geometric GNNs to expand the scope of their capabilities by incorporating mesh points alongside atoms and reimaging traditional mathematical operations in a trainable manner. Neural P$^3$M exhibits flexibility across a wide range of molecular systems and demonstrates remarkable accuracy in predicting energies and forces, outperforming on benchmarks such as the MD22 dataset. It also achieves an average improvement of 22% on the OE62 dataset while integrating with various architectures.

Shengnan An, Zexiong Ma, Zeqi Lin et al.

While many contemporary large language models (LLMs) can process lengthy input, they still struggle to fully utilize information within the long context, known as the lost-in-the-middle challenge. We hypothesize that it stems from insufficient explicit supervision during the long-context training, which fails to emphasize that any position in a long context can hold crucial information. Based on this intuition, our study presents information-intensive (IN2) training, a purely data-driven solution to overcome lost-in-the-middle. Specifically, IN2 training leverages a synthesized long-context question-answer dataset, where the answer requires (1) fine-grained information awareness on a short segment (~128 tokens) within a synthesized long context (4K-32K tokens), and (2) the integration and reasoning of information from two or more short segments. Through applying this information-intensive training on Mistral-7B, we present FILM-7B (FILl-in-the-Middle). To thoroughly assess the ability of FILM-7B for utilizing long contexts, we design three probing tasks that encompass various context styles (document, code, and structured-data context) and information retrieval patterns (forward, backward, and bi-directional retrieval). The probing results demonstrate that FILM-7B can robustly retrieve information from different positions in its 32K context window. Beyond these probing tasks, FILM-7B significantly improves the performance on real-world long-context tasks (e.g., 23.5->26.9 F1 score on NarrativeQA), while maintaining a comparable performance on short-context tasks (e.g., 59.3->59.2 accuracy on MMLU). Github Link: https://github.com/microsoft/FILM.

Miao Kang, Shengqi Wang, Sanping Zhou et al.

Motion forecasting plays a crucial role in autonomous driving, with the aim of predicting the future reasonable motions of traffic agents. Most existing methods mainly model the historical interactions between agents and the environment, and predict multi-modal trajectories in a feedforward process, ignoring potential trajectory changes caused by future interactions between agents. In this paper, we propose a novel Future Feedback Interaction Network (FFINet) to aggregate features the current observations and potential future interactions for trajectory prediction. Firstly, we employ different spatial-temporal encoders to embed the decomposed position vectors and the current position of each scene, providing rich features for the subsequent cross-temporal aggregation. Secondly, the relative interaction and cross-temporal aggregation strategies are sequentially adopted to integrate features in the current fusion module, observation interaction module, future feedback module and global fusion module, in which the future feedback module can enable the understanding of pre-action by feeding the influence of preview information to feedforward prediction. Thirdly, the comprehensive interaction features are further fed into final predictor to generate the joint predicted trajectories of multiple agents. Extensive experimental results show that our FFINet achieves the state-of-the-art performance on Argoverse 1 and Argoverse 2 motion forecasting benchmarks.

Haotian Wang, Meng Yang, Nanning Zheng

Monocular depth inference is a fundamental problem for scene perception of robots. Specific robots may be equipped with a camera plus an optional depth sensor of any type and located in various scenes of different scales, whereas recent advances derived multiple individual sub-tasks. It leads to additional burdens to fine-tune models for specific robots and thereby high-cost customization in large-scale industrialization. This paper investigates a unified task of monocular depth inference, which infers high-quality depth maps from all kinds of input raw data from various robots in unseen scenes. A basic benchmark G2-MonoDepth is developed for this task, which comprises four components: (a) a unified data representation RGB+X to accommodate RGB plus raw depth with diverse scene scale/semantics, depth sparsity ([0%, 100%]) and errors (holes/noises/blurs), (b) a novel unified loss to adapt to diverse depth sparsity/errors of input raw data and diverse scales of output scenes, (c) an improved network to well propagate diverse scene scales from input to output, and (d) a data augmentation pipeline to simulate all types of real artifacts in raw depth maps for training. G2-MonoDepth is applied in three sub-tasks including depth estimation, depth completion with different sparsity, and depth enhancement in unseen scenes, and it always outperforms SOTA baselines on both real-world data and synthetic data.

Han Wang, Jiayuan Zhang, Lipeng Wan et al.

Manipulation relationship detection (MRD) aims to guide the robot to grasp objects in the right order, which is important to ensure the safety and reliability of grasping in object stacked scenes. Previous works infer manipulation relationship by deep neural network trained with data collected from a predefined view, which has limitation in visual dislocation in unstructured environments. Multi-view data provide more comprehensive information in space, while a challenge of multi-view MRD is domain shift. In this paper, we propose a novel multi-view fusion framework, namely multi-view MRD network (MMRDN), which is trained by 2D and 3D multi-view data. We project the 2D data from different views into a common hidden space and fit the embeddings with a set of Von-Mises-Fisher distributions to learn the consistent representations. Besides, taking advantage of position information within the 3D data, we select a set of $K$ Maximum Vertical Neighbors (KMVN) points from the point cloud of each object pair, which encodes the relative position of these two objects. Finally, the features of multi-view 2D and 3D data are concatenated to predict the pairwise relationship of objects. Experimental results on the challenging REGRAD dataset show that MMRDN outperforms the state-of-the-art methods in multi-view MRD tasks. The results also demonstrate that our model trained by synthetic data is capable to transfer to real-world scenarios.

Yulong Liu, Yongqiang Ma, Guibo Zhu et al.

Deciphering visual content from functional Magnetic Resonance Imaging (fMRI) helps illuminate the human vision system. However, the scarcity of fMRI data and noise hamper brain decoding model performance. Previous approaches primarily employ subject-specific models, sensitive to training sample size. In this paper, we explore a straightforward but overlooked solution to address data scarcity. We propose shallow subject-specific adapters to map cross-subject fMRI data into unified representations. Subsequently, a shared deeper decoding model decodes cross-subject features into the target feature space. During training, we leverage both visual and textual supervision for multi-modal brain decoding. Our model integrates a high-level perception decoding pipeline and a pixel-wise reconstruction pipeline guided by high-level perceptions, simulating bottom-up and top-down processes in neuroscience. Empirical experiments demonstrate robust neural representation learning across subjects for both pipelines. Moreover, merging high-level and low-level information improves both low-level and high-level reconstruction metrics. Additionally, we successfully transfer learned general knowledge to new subjects by training new adapters with limited training data. Compared to previous state-of-the-art methods, notably pre-training-based methods (Mind-Vis and fMRI-PTE), our approach achieves comparable or superior results across diverse tasks, showing promise as an alternative method for cross-subject fMRI data pre-training. Our code and pre-trained weights will be publicly released at https://github.com/YulongBonjour/See_Through_Their_Minds.

Jingwen Fu, Nanning Zheng

This paper explores the generalization characteristics of iterative learning algorithms with bounded updates for non-convex loss functions, employing information-theoretic techniques. Our key contribution is a novel bound for the generalization error of these algorithms with bounded updates. Our approach introduces two main novelties: 1) we reformulate the mutual information as the uncertainty of updates, providing a new perspective, and 2) instead of using the chaining rule of mutual information, we employ a variance decomposition technique to decompose information across iterations, allowing for a simpler surrogate process. We analyze our generalization bound under various settings and demonstrate improved bounds. To bridge the gap between theory and practice, we also examine the previously observed scaling behavior in large language models. Ultimately, our work takes a further step for developing practical generalization theories.

Shitao Chen, Haolin Zhang, Nanning Zheng

3D object detection based on LiDAR point cloud and prior anchor boxes is a critical technology for autonomous driving environment perception and understanding. Nevertheless, an overlooked practical issue in existing methods is the ambiguity in training sample allocation based on box Intersection over Union (IoU_box). This problem impedes further enhancements in the performance of anchor-based LiDAR 3D object detectors. To tackle this challenge, this paper introduces a new training sample selection method that utilizes point cloud distribution for anchor sample quality measurement, named Point Assisted Sample Selection (PASS). This method has undergone rigorous evaluation on two widely utilized datasets. Experimental results demonstrate that the application of PASS elevates the average precision of anchor-based LiDAR 3D object detectors to a novel state-of-the-art, thereby proving the effectiveness of the proposed approach. The codes will be made available at https://github.com/XJTU-Haolin/Point_Assisted_Sample_Selection.

Sijie Liu, Ruisheng Su, Jianghang Su et al.

Accurate automated extraction of brain vessel centerlines from CTA images plays an important role in diagnosis and therapy of cerebrovascular diseases, such as stroke. However, this task remains challenging due to the complex cerebrovascular structure, the varying imaging quality, and vessel pathology effects. In this paper, we consider automatic lumen segmentation generation without additional annotation effort by physicians and more effective use of the generated lumen segmentation for improved centerline extraction performance. We propose an automated framework for brain vessel centerline extraction from CTA images. The framework consists of four major components: (1) pre-processing approaches that register CTA images with a CT atlas and divide these images into input patches, (2) lumen segmentation generation from annotated vessel centerlines using graph cuts and robust kernel regression, (3) a dual-branch topology-aware UNet (DTUNet) that can effectively utilize the annotated vessel centerlines and the generated lumen segmentation through a topology-aware loss (TAL) and its dual-branch design, and (4) post-processing approaches that skeletonize the predicted lumen segmentation. Extensive experiments on a multi-center dataset demonstrate that the proposed framework outperforms state-of-the-art methods in terms of average symmetric centerline distance (ASCD) and overlap (OV). Subgroup analyses further suggest that the proposed framework holds promise in clinical applications for stroke treatment. Code is publicly available at https://github.com/Liusj-gh/DTUNet.

Chaoyun Wang, Quanxin Huang, I-Chao Shen et al.

Document rectification in real-world scenarios poses significant challenges due to extreme variations in camera perspectives and physical distortions. Driven by the insight that complex transformations can be decomposed and resolved progressively, we introduce a novel multi-stage framework that progressively reverses distinct distortion types in a coarse-to-fine manner. Specifically, our framework first performs a global affine transformation to correct perspective distortions arising from the camera's viewpoint, then rectifies geometric deformations resulting from physical paper curling and folding, and finally employs a content-aware iterative process to eliminate fine-grained content distortions. To address limitations in existing evaluation protocols, we also propose two enhanced metrics: layout-aligned OCR metrics (AED/ACER) for a stable assessment that decouples geometric rectification quality from the layout analysis errors of OCR engines, and masked AD/AAD (AD-M/AAD-M) tailored for accurately evaluating geometric distortions in documents with incomplete boundaries. Extensive experiments show that our method establishes new state-of-the-art performance on multiple challenging benchmarks, yielding a substantial reduction of 14.1\%--34.7\% in the AAD metric and demonstrating superior efficacy in real-world applications. The code will be publicly available at https://github.com/chaoyunwang/ArbDR.

Chen Li, Weiqi Wang, Jingcheng Hu et al.

Mathematical capabilities were previously believed to emerge in common language models only at a very large scale or require extensive math-related pre-training. This paper shows that the LLaMA-2 7B model with common pre-training already exhibits strong mathematical abilities, as evidenced by its impressive accuracy of 97.7% and 72.0% on the GSM8K and MATH benchmarks, respectively, when selecting the best response from 256 random generations. The primary issue with the current base model is the difficulty in consistently eliciting its inherent mathematical capabilities. Notably, the accuracy for the first answer drops to 49.5% and 7.9% on the GSM8K and MATH benchmarks, respectively. We find that simply scaling up the SFT data can significantly enhance the reliability of generating correct answers. However, the potential for extensive scaling is constrained by the scarcity of publicly available math questions. To overcome this limitation, we employ synthetic data, which proves to be nearly as effective as real data and shows no clear saturation when scaled up to approximately one million samples. This straightforward approach achieves an accuracy of 82.6% on GSM8K and 40.6% on MATH using LLaMA-2 7B models, surpassing previous models by 14.2% and 20.8%, respectively. We also provide insights into scaling behaviors across different reasoning complexities and error types.

Zijian Kang, Peizhen Zhang, Xiangyu Zhang et al.

Knowledge distillation has shown great success in classification, however, it is still challenging for detection. In a typical image for detection, representations from different locations may have different contributions to detection targets, making the distillation hard to balance. In this paper, we propose a conditional distillation framework to distill the desired knowledge, namely knowledge that is beneficial in terms of both classification and localization for every instance. The framework introduces a learnable conditional decoding module, which retrieves information given each target instance as query. Specifically, we encode the condition information as query and use the teacher's representations as key. The attention between query and key is used to measure the contribution of different features, guided by a localization-recognition-sensitive auxiliary task. Extensive experiments demonstrate the efficacy of our method: we observe impressive improvements under various settings. Notably, we boost RetinaNet with ResNet-50 backbone from 37.4 to 40.7 mAP (+3.3) under 1x schedule, that even surpasses the teacher (40.4 mAP) with ResNet-101 backbone under 3x schedule. Code has been released on https://github.com/megvii-research/ICD.

Peizhen Zhang, Zijian Kang, Tong Yang et al.

In this paper, we propose the first self-distillation framework for general object detection, termed LGD (Label-Guided self-Distillation). Previous studies rely on a strong pretrained teacher to provide instructive knowledge that could be unavailable in real-world scenarios. Instead, we generate an instructive knowledge based only on student representations and regular labels. Our framework includes sparse label-appearance encoder, inter-object relation adapter and intra-object knowledge mapper that jointly form an implicit teacher at training phase, dynamically dependent on labels and evolving student representations. They are trained end-to-end with detector and discarded in inference. Experimentally, LGD obtains decent results on various detectors, datasets, and extensive tasks like instance segmentation. For example in MS-COCO dataset, LGD improves RetinaNet with ResNet-50 under 2x single-scale training from 36.2% to 39.0% mAP (+ 2.8%). It boosts much stronger detectors like FCOS with ResNeXt-101 DCN v2 under 2x multi-scale training from 46.1% to 47.9% (+ 1.8%). Compared with a classical teacher-based method FGFI, LGD not only performs better without requiring pretrained teacher but also reduces 51% training cost beyond inherent student learning. Codes are available at https://github.com/megvii-research/LGD.

Zixin Zhu, Wei Tang, Le Wang et al.

Effectively tackling the problem of temporal action localization (TAL) necessitates a visual representation that jointly pursues two confounding goals, i.e., fine-grained discrimination for temporal localization and sufficient visual invariance for action classification. We address this challenge by enriching both the local and global contexts in the popular two-stage temporal localization framework, where action proposals are first generated followed by action classification and temporal boundary regression. Our proposed model, dubbed ContextLoc, can be divided into three sub-networks: L-Net, G-Net and P-Net. L-Net enriches the local context via fine-grained modeling of snippet-level features, which is formulated as a query-and-retrieval process. G-Net enriches the global context via higher-level modeling of the video-level representation. In addition, we introduce a novel context adaptation module to adapt the global context to different proposals. P-Net further models the context-aware inter-proposal relations. We explore two existing models to be the P-Net in our experiments. The efficacy of our proposed method is validated by experimental results on the THUMOS14 (54.3\% at tIoU@0.5) and ActivityNet v1.3 (56.01\% at tIoU@0.5) datasets, which outperforms recent states of the art. Code is available at https://github.com/buxiangzhiren/ContextLoc.