Zhirong Wu

Yutong Chen, Fangyun Wei, Xiao Sun et al. · tsinghua

This paper proposes a simple transfer learning baseline for sign language translation. Existing sign language datasets (e.g. PHOENIX-2014T, CSL-Daily) contain only about 10K-20K pairs of sign videos, gloss annotations and texts, which are an order of magnitude smaller than typical parallel data for training spoken language translation models. Data is thus a bottleneck for training effective sign language translation models. To mitigate this problem, we propose to progressively pretrain the model from general-domain datasets that include a large amount of external supervision to within-domain datasets. Concretely, we pretrain the sign-to-gloss visual network on the general domain of human actions and the within-domain of a sign-to-gloss dataset, and pretrain the gloss-to-text translation network on the general domain of a multilingual corpus and the within-domain of a gloss-to-text corpus. The joint model is fine-tuned with an additional module named the visual-language mapper that connects the two networks. This simple baseline surpasses the previous state-of-the-art results on two sign language translation benchmarks, demonstrating the effectiveness of transfer learning. With its simplicity and strong performance, this approach can serve as a solid baseline for future research. Code and models are available at: https://github.com/FangyunWei/SLRT.

Zhihang Zhong, Xiao Sun, Zhirong Wu et al.

We study the challenging problem of recovering detailed motion from a single motion-blurred image. Existing solutions to this problem estimate a single image sequence without considering the motion ambiguity for each region. Therefore, the results tend to converge to the mean of the multi-modal possibilities. In this paper, we explicitly account for such motion ambiguity, allowing us to generate multiple plausible solutions all in sharp detail. The key idea is to introduce a motion guidance representation, which is a compact quantization of 2D optical flow with only four discrete motion directions. Conditioned on the motion guidance, the blur decomposition is led to a specific, unambiguous solution by using a novel two-stage decomposition network. We propose a unified framework for blur decomposition, which supports various interfaces for generating our motion guidance, including human input, motion information from adjacent video frames, and learning from a video dataset. Extensive experiments on synthesized datasets and real-world data show that the proposed framework is qualitatively and quantitatively superior to previous methods, and also offers the merit of producing physically plausible and diverse solutions. Code is available at https://github.com/zzh-tech/Animation-from-Blur.

Zhihang Zhong, Mingdeng Cao, Xiao Sun et al.

Rolling shutter (RS) distortion can be interpreted as the result of picking a row of pixels from instant global shutter (GS) frames over time during the exposure of the RS camera. This means that the information of each instant GS frame is partially, yet sequentially, embedded into the row-dependent distortion. Inspired by this fact, we address the challenging task of reversing this process, i.e., extracting undistorted GS frames from images suffering from RS distortion. However, since RS distortion is coupled with other factors such as readout settings and the relative velocity of scene elements to the camera, models that only exploit the geometric correlation between temporally adjacent images suffer from poor generality in processing data with different readout settings and dynamic scenes with both camera motion and object motion. In this paper, instead of two consecutive frames, we propose to exploit a pair of images captured by dual RS cameras with reversed RS directions for this highly challenging task. Grounded on the symmetric and complementary nature of dual reversed distortion, we develop a novel end-to-end model, IFED, to generate dual optical flow sequence through iterative learning of the velocity field during the RS time. Extensive experimental results demonstrate that IFED is superior to naive cascade schemes, as well as the state-of-the-art which utilizes adjacent RS images. Most importantly, although it is trained on a synthetic dataset, IFED is shown to be effective at retrieving GS frame sequences from real-world RS distorted images of dynamic scenes. Code is available at https://github.com/zzh-tech/Dual-Reversed-RS.

Huimin Wu, Chenyang Lei, Xiao Sun et al.

Self-supervised representation learning follows a paradigm of withholding some part of the data and tasking the network to predict it from the remaining part. Among many techniques, data augmentation lies at the core for creating the information gap. Towards this end, masking has emerged as a generic and powerful tool where content is withheld along the sequential dimension, e.g., spatial in images, temporal in audio, and syntactic in language. In this paper, we explore the orthogonal channel dimension for generic data augmentation by exploiting precision redundancy. The data for each channel is quantized through a non-uniform quantizer, with the quantized value sampled randomly within randomly sampled quantization bins. From another perspective, quantization is analogous to channel-wise masking, as it removes the information within each bin, but preserves the information across bins. Our approach significantly surpasses existing generic data augmentation methods, while showing on par performance against modality-specific augmentations. We comprehensively evaluate our approach on vision, audio, 3D point clouds, as well as the DABS benchmark which is comprised of various data modalities. The code is available at https: //github.com/microsoft/random_quantize.

Chenguo Lin, Xumeng Wen, Wei Cao et al.

Recent research on time-series self-supervised models shows great promise in learning semantic representations. However, it has been limited to small-scale datasets, e.g., thousands of temporal sequences. In this work, we make key technical contributions that are tailored to the numerical properties of time-series data and allow the model to scale to large datasets, e.g., millions of temporal sequences. We adopt the Transformer architecture by first partitioning the input into non-overlapping windows. Each window is then characterized by its normalized shape and two scalar values denoting the mean and standard deviation within each window. To embed scalar values that may possess arbitrary numerical amplitudes in a high-dimensional space, we propose a numerically multi-scaled embedding module enumerating all possible numerical scales for the scalars. The model undergoes pretraining with a simple contrastive objective on a large-scale dataset over a million sequences collected by merging existing public data. We study its transfer performance on a number of univariate and multivariate classification tasks, few shot learning, unsupervised clustering and anomaly detection benchmarks. Our method exhibits remarkable improvement against previous pretraining approaches and establishes the new state of the art, even compared with domain-specific non-learning-based methods. Code is available at: \url{https://github.com/chenguolin/NuTime}.

Long Lian, Zhirong Wu, Stella X. Yu

We study learning object segmentation from unlabeled videos. Humans can easily segment moving objects without knowing what they are. The Gestalt law of common fate, i.e., what move at the same speed belong together, has inspired unsupervised object discovery based on motion segmentation. However, common fate is not a reliable indicator of objectness: Parts of an articulated / deformable object may not move at the same speed, whereas shadows / reflections of an object always move with it but are not part of it. Our insight is to bootstrap objectness by first learning image features from relaxed common fate and then refining them based on visual appearance grouping within the image itself and across images statistically. Specifically, we learn an image segmenter first in the loop of approximating optical flow with constant segment flow plus small within-segment residual flow, and then by refining it for more coherent appearance and statistical figure-ground relevance. On unsupervised video object segmentation, using only ResNet and convolutional heads, our model surpasses the state-of-the-art by absolute gains of 7/9/5% on DAVIS16 / STv2 / FBMS59 respectively, demonstrating the effectiveness of our ideas. Our code is publicly available.

Zhihang Zhong, Mingxi Cheng, Zhirong Wu et al.

Image cropping has progressed tremendously under the data-driven paradigm. However, current approaches do not account for the intentions of the user, which is an issue especially when the composition of the input image is complex. Moreover, labeling of cropping data is costly and hence the amount of data is limited, leading to poor generalization performance of current algorithms in the wild. In this work, we take advantage of vision-language models as a foundation for creating robust and user-intentional cropping algorithms. By adapting a transformer decoder with a pre-trained CLIP-based detection model, OWL-ViT, we develop a method to perform cropping with a text or image query that reflects the user's intention as guidance. In addition, our pipeline design allows the model to learn text-conditioned aesthetic cropping with a small cropping dataset, while inheriting the open-vocabulary ability acquired from millions of text-image pairs. We validate our model through extensive experiments on existing datasets as well as a new cropping test set we compiled that is characterized by content ambiguity.

Zhirong Wu, Zihang Lai, Xiao Sun et al.

The paper presents a scalable approach for learning spatially distributed visual representations over individual tokens and a holistic instance representation simultaneously. We use self-attention blocks to represent spatially distributed tokens, followed by cross-attention blocks to aggregate the holistic image instance. The core of the approach is the use of extremely large token masking (75\%-90\%) as the data augmentation for supervision. Our model, named ExtreMA, follows the plain BYOL approach where the instance representation from the unmasked subset is trained to predict that from the intact input. Instead of encouraging invariance across inputs, the model is required to capture informative variations in an image. The paper makes three contributions: 1) It presents random masking as a strong and computationally efficient data augmentation for siamese representation learning. 2) With multiple sampling per instance, extreme masking greatly speeds up learning and improves performance with more data. 3) ExtreMA obtains stronger linear probing performance than masked modeling methods, and better transfer performance than prior contrastive models.

Long Lian, Zhirong Wu, Stella X. Yu

We present IMAS, a method that segments the primary objects in videos without manual annotation in training or inference. Previous methods in unsupervised video object segmentation (UVOS) have demonstrated the effectiveness of motion as either input or supervision for segmentation. However, motion signals may be uninformative or even misleading in cases such as deformable objects and objects with reflections, causing unsatisfactory segmentation. In contrast, IMAS achieves Improved UVOS with Motion-Appearance Synergy. Our method has two training stages: 1) a motion-supervised object discovery stage that deals with motion-appearance conflicts through a learnable residual pathway; 2) a refinement stage with both low- and high-level appearance supervision to correct model misconceptions learned from misleading motion cues. Additionally, we propose motion-semantic alignment as a model-agnostic annotation-free hyperparam tuning method. We demonstrate its effectiveness in tuning critical hyperparams previously tuned with human annotation or hand-crafted hyperparam-specific metrics. IMAS greatly improves the segmentation quality on several common UVOS benchmarks. For example, we surpass previous methods by 8.3% on DAVIS16 benchmark with only standard ResNet and convolutional heads. We intend to release our code for future research and applications.

Jialiang Zhu, Gongrui Zhang, Xiaolong Ma et al.

LLM-based deep research agents are largely built on the ReAct framework. This linear design makes it difficult to revisit earlier states, branch into alternative search directions, or maintain global awareness under long contexts, often leading to local optima, redundant exploration, and inefficient search. We propose Re-TRAC, an agentic framework that performs cross-trajectory exploration by generating a structured state representation after each trajectory to summarize evidence, uncertainties, failures, and future plans, and conditioning subsequent trajectories on this state representation. This enables iterative reflection and globally informed planning, reframing research as a progressive process. Empirical results show that Re-TRAC consistently outperforms ReAct by 15-20% on BrowseComp with frontier LLMs. For smaller models, we introduce Re-TRAC-aware supervised fine-tuning, achieving state-of-the-art performance at comparable scales. Notably, Re-TRAC shows a monotonic reduction in tool calls and token usage across rounds, indicating progressively targeted exploration driven by cross-trajectory reflection rather than redundant search.

Yuwei Sun, Hideya Ochiai, Zhirong Wu et al.

Emerging from the pairwise attention in conventional Transformers, there is a growing interest in sparse attention mechanisms that align more closely with localized, contextual learning in the biological brain. Existing studies such as the Coordination method employ iterative cross-attention mechanisms with a bottleneck to enable the sparse association of inputs. However, these methods are parameter inefficient and fail in more complex relational reasoning tasks. To this end, we propose Associative Transformer (AiT) to enhance the association among sparsely attended input tokens, improving parameter efficiency and performance in various vision tasks such as classification and relational reasoning. AiT leverages a learnable explicit memory comprising specialized priors that guide bottleneck attentions to facilitate the extraction of diverse localized tokens. Moreover, AiT employs an associative memory-based token reconstruction using a Hopfield energy function. The extensive empirical experiments demonstrate that AiT requires significantly fewer parameters and attention layers outperforming a broad range of sparse Transformer models. Additionally, AiT outperforms the SOTA sparse Transformer models including the Coordination method on the Sort-of-CLEVR dataset.

Xiaodong Wang, Langling Huang, Zhirong Wu et al.

The development of multimodal large language models (MLLMs) has advanced general video understanding. However, existing video evaluation benchmarks primarily focus on non-interactive videos, such as movies and recordings. To fill this gap, this paper proposes the first omnimodal benchmark for interactive livestream videos, LiViBench. It features a diverse set of 24 tasks, highlighting the perceptual, reasoning, and livestream-specific challenges. To efficiently construct the dataset, we design a standardized semi-automatic annotation workflow that incorporates the human-in-the-loop at multiple stages. The workflow leverages multiple MLLMs to form a multi-agent system for comprehensive video description and uses a seed-question-driven method to construct high-quality annotations. All interactive videos in the benchmark include audio, speech, and real-time comments modalities. To enhance models' understanding of interactive videos, we design tailored two-stage instruction-tuning and propose a Video-to-Comment Retrieval (VCR) module to improve the model's ability to utilize real-time comments. Based on these advancements, we develop LiVi-LLM-7B, an MLLM with enhanced knowledge of interactive livestreams. Experiments show that our model outperforms larger open-source models with up to 72B parameters, narrows the gap with leading proprietary models on LiViBench, and achieves enhanced performance on general video benchmarks, including VideoMME, LongVideoBench, MLVU, and VideoEval-Pro.

Gongrui Zhang, Jialiang Zhu, Ruiqi Yang et al. · microsoft-research

Building Large Language Model agents that expand their capabilities by interacting with external tools represents a new frontier in AI research and applications. In this paper, we introduce InfoAgent, a deep research agent powered by an innovative data synthesis pipeline and orchestrated web search tools. To construct challenging, hard-to-find queries,we build entity trees and apply sub-tree sampling with entity fuzzification to systematically increase question difficulty. Unlike prior work that relies heavily on commercial search tools, we develop a dedicated self-hosted search infrastructure, enhancing transparency of agent environments and facilitating further advancement of agent capacity. We evaluate the effectiveness of our data pipeline by measuring the average number of tool calls required to correctly answer a question, and also show that our agent yields better performance when equipped with our tools. Our \mbox{InfoAgent} is post-trained from Qwen3-14B using a two-stage recipe: cold-start supervised finetuning to instill long-horizon search behaviors, followed by reinforcement learning which significantly improves reasoning-driven tool use. With our methods, InfoAgent achieves 15.3\% accuracy on BrowseComp, 29.2\% on BrowseComp-ZH, and 40.4\% on Xbench-DS, outperforming prior open-source deep research agents such as WebSailor-72B and DeepDive-32B.

Huimin Wu, Kwang-Ting Cheng, Stephen Lin et al.

This paper presents an investigation of vision transformer learning for multi-view geometry tasks, such as optical flow estimation, by fine-tuning video foundation models. Unlike previous methods that involve custom architectural designs and task-specific pretraining, our research finds that general-purpose models pretrained on videos can be readily transferred to multi-view problems with minimal adaptation. The core insight is that general-purpose attention between patches learns temporal and spatial information for geometric reasoning. We demonstrate that appending a linear decoder to the Transformer backbone produces satisfactory results, and iterative refinement can further elevate performance to stateof-the-art levels. This conceptually simple approach achieves top cross-dataset generalization results for optical flow estimation with end-point error (EPE) of 0.69, 1.78, and 3.15 on the Sintel clean, Sintel final, and KITTI datasets, respectively. Our method additionally establishes a new record on the online test benchmark with EPE values of 0.79, 1.88, and F1 value of 3.79. Applications to 3D depth estimation and stereo matching also show strong performance, illustrating the versatility of video-pretrained models in addressing geometric vision tasks.

Tian Xie, Zitian Gao, Qingnan Ren et al.

Inspired by the success of DeepSeek-R1, we explore the potential of rule-based reinforcement learning (RL) in large reasoning models. To analyze reasoning dynamics, we use synthetic logic puzzles as training data due to their controllable complexity and straightforward answer verification. We make some key technical contributions that lead to effective and stable RL training: a system prompt that emphasizes the thinking and answering process, a stringent format reward function that penalizes outputs for taking shortcuts, and a straightforward training recipe that achieves stable convergence. Our 7B model develops advanced reasoning skills-such as reflection, verification, and summarization-that are absent from the logic corpus. Remarkably, after training on just 5K logic problems, it demonstrates generalization abilities to the challenging math benchmarks AIME and AMC.

Xudong Wang, Zhirong Wu, Long Lian et al.

Pseudo-labels are confident predictions made on unlabeled target data by a classifier trained on labeled source data. They are widely used for adapting a model to unlabeled data, e.g., in a semi-supervised learning setting. Our key insight is that pseudo-labels are naturally imbalanced due to intrinsic data similarity, even when a model is trained on balanced source data and evaluated on balanced target data. If we address this previously unknown imbalanced classification problem arising from pseudo-labels instead of ground-truth training labels, we could remove model biases towards false majorities created by pseudo-labels. We propose a novel and effective debiased learning method with pseudo-labels, based on counterfactual reasoning and adaptive margins: The former removes the classifier response bias, whereas the latter adjusts the margin of each class according to the imbalance of pseudo-labels. Validated by extensive experimentation, our simple debiased learning delivers significant accuracy gains over the state-of-the-art on ImageNet-1K: 26% for semi-supervised learning with 0.2% annotations and 9% for zero-shot learning. Our code is available at: https://github.com/frank-xwang/debiased-pseudo-labeling.

Fangyun Wei, Yue Gao, Zhirong Wu et al.

Image-level contrastive representation learning has proven to be highly effective as a generic model for transfer learning. Such generality for transfer learning, however, sacrifices specificity if we are interested in a certain downstream task. We argue that this could be sub-optimal and thus advocate a design principle which encourages alignment between the self-supervised pretext task and the downstream task. In this paper, we follow this principle with a pretraining method specifically designed for the task of object detection. We attain alignment in the following three aspects: 1) object-level representations are introduced via selective search bounding boxes as object proposals; 2) the pretraining network architecture incorporates the same dedicated modules used in the detection pipeline (e.g. FPN); 3) the pretraining is equipped with object detection properties such as object-level translation invariance and scale invariance. Our method, called Selective Object COntrastive learning (SoCo), achieves state-of-the-art results for transfer performance on COCO detection using a Mask R-CNN framework. Code is available at https://github.com/hologerry/SoCo.

Yizhuo Zhang, Zhirong Wu, Houwen Peng et al.

Semi-supervised video object segmentation aims to separate a target object from a video sequence, given the mask in the first frame. Most of current prevailing methods utilize information from additional modules trained in other domains like optical flow and instance segmentation, and as a result they do not compete with other methods on common ground. To address this issue, we propose a simple yet strong transductive method, in which additional modules, datasets, and dedicated architectural designs are not needed. Our method takes a label propagation approach where pixel labels are passed forward based on feature similarity in an embedding space. Different from other propagation methods, ours diffuses temporal information in a holistic manner which take accounts of long-term object appearance. In addition, our method requires few additional computational overhead, and runs at a fast $\sim$37 fps speed. Our single model with a vanilla ResNet50 backbone achieves an overall score of 72.3 on the DAVIS 2017 validation set and 63.1 on the test set. This simple yet high performing and efficient method can serve as a solid baseline that facilitates future research. Code and models are available at \url{https://github.com/microsoft/transductive-vos.pytorch}.

Xumeng Wen, Zihan Liu, Shun Zheng et al.

Recent advancements in long chain-of-thought (CoT) reasoning, particularly through the Group Relative Policy Optimization algorithm used by DeepSeek-R1, have led to significant interest in the potential of Reinforcement Learning with Verifiable Rewards (RLVR) for Large Language Models (LLMs). While RLVR promises to improve reasoning by allowing models to learn from free exploration, there remains debate over whether it truly enhances reasoning abilities or simply boosts sampling efficiency. This paper systematically investigates the impact of RLVR on LLM reasoning. We revisit Pass@K experiments and demonstrate that RLVR can extend the reasoning boundary for both mathematical and coding tasks. This is supported by our introduction of a novel evaluation metric, CoT-Pass@K, which captures reasoning success by accounting for both the final answer and intermediate reasoning steps. Furthermore, we present a theoretical framework explaining RLVR's incentive mechanism, demonstrating how it can encourage correct reasoning even when rewards are based solely on answer correctness. Our analysis of RLVR's training dynamics reveals that it incentivizes correct reasoning early in the process, with substantial improvements in reasoning quality confirmed through extensive evaluations. These findings provide strong evidence of RLVR's potential to enhance LLM reasoning, offering valuable insights into its mechanisms and performance improvements.

Xiaodong Wang, Zhirong Wu, Peixi Peng · pku

Driving world models are used to simulate futures by video generation based on the condition of the current state and actions. However, current models often suffer serious error accumulations when predicting the long-term future, which limits the practical application. Recent studies utilize the Diffusion Transformer (DiT) as the backbone of driving world models to improve learning flexibility. However, these models are always trained on short video clips (high fps and short duration), and multiple roll-out generations struggle to produce consistent and reasonable long videos due to the training-inference gap. To this end, we propose several solutions to build a simple yet effective long-term driving world model. First, we hierarchically decouple world model learning into large motion learning and bidirectional continuous motion learning. Then, considering the continuity of driving scenes, we propose a simple distillation method where fine-grained video flows are self-supervised signals for coarse-grained flows. The distillation is designed to improve the coherence of infinite video generation. The coarse-grained and fine-grained modules are coordinated to generate long-term and temporally coherent videos. In the public benchmark NuScenes, compared with the state-of-the-art front-view model, our model improves FVD by $27\%$ and reduces inference time by $85\%$ for the video task of generating 110+ frames. More videos (including 90s duration) are available at https://Wang-Xiaodong1899.github.io/longdwm/.

Kai Qiu, Huishuai Zhang, Zhirong Wu et al.

Training foundation models on extensive datasets and then finetuning them on specific tasks has emerged as the mainstream approach in artificial intelligence. However, the model robustness, which is a critical aspect for safety, is often optimized for each specific task rather than at the pretraining stage. In this paper, we propose a method for pretraining certifiably robust models that can be readily finetuned for adaptation to a particular task. A key challenge is dealing with the compromise between semantic learning and robustness. We address this with a simple yet highly effective strategy based on significantly broadening the pretraining data distribution, which is shown to greatly benefit finetuning for downstream tasks. Through pretraining on a mixture of clean and various noisy images, we find that surprisingly strong certified accuracy can be achieved even when finetuning on only clean images. Furthermore, this strategy requires just a single model to deal with various noise levels, thus substantially reducing computational costs in relation to previous works that employ multiple models. Despite using just one model, our method can still yield results that are on par with, or even superior to, existing multi-model methods.

Kenneth Li, Xiao Sun, Zhirong Wu et al.

For human action understanding, a popular research direction is to analyze short video clips with unambiguous semantic content, such as jumping and drinking. However, methods for understanding short semantic actions cannot be directly translated to long human dynamics such as dancing, where it becomes challenging even to label the human movements semantically. Meanwhile, the natural language processing (NLP) community has made progress in solving a similar challenge of annotation scarcity by large-scale pre-training, which improves several downstream tasks with one model. In this work, we study how to segment and cluster videos into recurring temporal patterns in a self-supervised way, namely acton discovery, the main roadblock towards video tokenization. We propose a two-stage framework that first obtains a frame-wise representation by contrasting two augmented views of video frames conditioned on their temporal context. The frame-wise representations across a collection of videos are then clustered by K-means. Actons are then automatically extracted by forming a continuous motion sequence from frames within the same cluster. We evaluate the frame-wise representation learning step by Kendall's Tau and the lexicon building step by normalized mutual information and language entropy. We also study three applications of this tokenization: genre classification, action segmentation, and action composition. On the AIST++ and PKU-MMD datasets, actons bring significant performance improvements compared to several baselines.

Ceyuan Yang, Yujun Shen, Zhiyi Zhang et al.

This work aims at transferring a Generative Adversarial Network (GAN) pre-trained on one image domain to a new domain referring to as few as just one target image. The main challenge is that, under limited supervision, it is extremely difficult to synthesize photo-realistic and highly diverse images, while acquiring representative characters of the target. Different from existing approaches that adopt the vanilla fine-tuning strategy, we import two lightweight modules to the generator and the discriminator respectively. Concretely, we introduce an attribute adaptor into the generator yet freeze its original parameters, through which it can reuse the prior knowledge to the most extent and hence maintain the synthesis quality and diversity. We then equip the well-learned discriminator backbone with an attribute classifier to ensure that the generator captures the appropriate characters from the reference. Furthermore, considering the poor diversity of the training data (i.e., as few as only one image), we propose to also constrain the diversity of the generative domain in the training process, alleviating the optimization difficulty. Our approach brings appealing results under various settings, substantially surpassing state-of-the-art alternatives, especially in terms of synthesis diversity. Noticeably, our method works well even with large domain gaps, and robustly converges within a few minutes for each experiment.

Runtao Liu, Zhirong Wu, Stella X. Yu et al.

Humans can easily segment moving objects without knowing what they are. That objectness could emerge from continuous visual observations motivates us to model grouping and movement concurrently from unlabeled videos. Our premise is that a video has different views of the same scene related by moving components, and the right region segmentation and region flow would allow mutual view synthesis which can be checked from the data itself without any external supervision. Our model starts with two separate pathways: an appearance pathway that outputs feature-based region segmentation for a single image, and a motion pathway that outputs motion features for a pair of images. It then binds them in a conjoint representation called segment flow that pools flow offsets over each region and provides a gross characterization of moving regions for the entire scene. By training the model to minimize view synthesis errors based on segment flow, our appearance and motion pathways learn region segmentation and flow estimation automatically without building them up from low-level edges or optical flows respectively. Our model demonstrates the surprising emergence of objectness in the appearance pathway, surpassing prior works on zero-shot object segmentation from an image, moving object segmentation from a video with unsupervised test-time adaptation, and semantic image segmentation by supervised fine-tuning. Our work is the first truly end-to-end zero-shot object segmentation from videos. It not only develops generic objectness for segmentation and tracking, but also outperforms prevalent image-based contrastive learning methods without augmentation engineering.

Ceyuan Yang, Zhirong Wu, Bolei Zhou et al.

Prior research on self-supervised learning has led to considerable progress on image classification, but often with degraded transfer performance on object detection. The objective of this paper is to advance self-supervised pretrained models specifically for object detection. Based on the inherent difference between classification and detection, we propose a new self-supervised pretext task, called instance localization. Image instances are pasted at various locations and scales onto background images. The pretext task is to predict the instance category given the composited images as well as the foreground bounding boxes. We show that integration of bounding boxes into pretraining promotes better task alignment and architecture alignment for transfer learning. In addition, we propose an augmentation method on the bounding boxes to further enhance the feature alignment. As a result, our model becomes weaker at Imagenet semantic classification but stronger at image patch localization, with an overall stronger pretrained model for object detection. Experimental results demonstrate that our approach yields state-of-the-art transfer learning results for object detection on PASCAL VOC and MSCOCO.

Peng-Shuai Wang, Yu-Qi Yang, Qian-Fang Zou et al.

Although unsupervised feature learning has demonstrated its advantages to reducing the workload of data labeling and network design in many fields, existing unsupervised 3D learning methods still cannot offer a generic network for various shape analysis tasks with competitive performance to supervised methods. In this paper, we propose an unsupervised method for learning a generic and efficient shape encoding network for different shape analysis tasks. The key idea of our method is to jointly encode and learn shape and point features from unlabeled 3D point clouds. For this purpose, we adapt HR-Net to octree-based convolutional neural networks for jointly encoding shape and point features with fused multiresolution subnetworks and design a simple-yet-efficient Multiresolution Instance Discrimination (MID) loss for jointly learning the shape and point features. Our network takes a 3D point cloud as input and output both shape and point features. After training, the network is concatenated with simple task-specific back-end layers and fine-tuned for different shape analysis tasks. We evaluate the efficacy and generality of our method and validate our network and loss design with a set of shape analysis tasks, including shape classification, semantic shape segmentation, as well as shape registration tasks. With simple back-ends, our network demonstrates the best performance among all unsupervised methods and achieves competitive performance to supervised methods, especially in tasks with a small labeled dataset. For fine-grained shape segmentation, our method even surpasses existing supervised methods by a large margin.

Nanxuan Zhao, Zhirong Wu, Rynson W. H. Lau et al.

Contrastive visual pretraining based on the instance discrimination pretext task has made significant progress. Notably, recent work on unsupervised pretraining has shown to surpass the supervised counterpart for finetuning downstream applications such as object detection and segmentation. It comes as a surprise that image annotations would be better left unused for transfer learning. In this work, we investigate the following problems: What makes instance discrimination pretraining good for transfer learning? What knowledge is actually learned and transferred from these models? From this understanding of instance discrimination, how can we better exploit human annotation labels for pretraining? Our findings are threefold. First, what truly matters for the transfer is low-level and mid-level representations, not high-level representations. Second, the intra-category invariance enforced by the traditional supervised model weakens transferability by increasing task misalignment. Finally, supervised pretraining can be strengthened by following an exemplar-based approach without explicit constraints among the instances within the same category.

Nanxuan Zhao, Zhirong Wu, Rynson W. H. Lau et al.

Recent progress in contrastive learning has revolutionized unsupervised representation learning. Concretely, multiple views (augmentations) from the same image are encouraged to map to the similar embeddings, while views from different images are pulled apart. In this paper, through visualizing and diagnosing classification errors, we observe that current contrastive models are ineffective at localizing the foreground object, limiting their ability to extract discriminative high-level features. This is due to the fact that view generation process considers pixels in an image uniformly. To address this problem, we propose a data-driven approach for learning invariance to backgrounds. It first estimates foreground saliency in images and then creates augmentations by copy-and-pasting the foreground onto a variety of backgrounds. The learning still follows the instance discrimination pretext task, so that the representation is trained to disregard background content and focus on the foreground. We study a variety of saliency estimation methods, and find that most methods lead to improvements for contrastive learning. With this approach (DiLo), significant performance is achieved for self-supervised learning on ImageNet classification, and also for object detection on PASCAL VOC and MSCOCO.

Bin Liu, Zhirong Wu, Han Hu et al.

We study object recognition under the constraint that each object class is only represented by very few observations. Semi-supervised learning, transfer learning, and few-shot recognition all concern with achieving fast generalization with few labeled data. In this paper, we propose a generic framework that utilizes unlabeled data to aid generalization for all three tasks. Our approach is to create much more training data through label propagation from the few labeled examples to a vast collection of unannotated images. The main contribution of the paper is that we show such a label propagation scheme can be highly effective when the similarity metric used for propagation is transferred from other related domains. We test various combinations of supervised and unsupervised metric learning methods with various label propagation algorithms. We find that our framework is very generic without being sensitive to any specific techniques. By taking advantage of unlabeled data in this way, we achieve significant improvements on all three tasks.

Zhirong Wu, Alexei A. Efros, Stella X. Yu

Current major approaches to visual recognition follow an end-to-end formulation that classifies an input image into one of the pre-determined set of semantic categories. Parametric softmax classifiers are a common choice for such a closed world with fixed categories, especially when big labeled data is available during training. However, this becomes problematic for open-set scenarios where new categories are encountered with very few examples for learning a generalizable parametric classifier. We adopt a non-parametric approach for visual recognition by optimizing feature embeddings instead of parametric classifiers. We use a deep neural network to learn the visual feature that preserves the neighborhood structure in the semantic space, based on the Neighborhood Component Analysis (NCA) criterion. Limited by its computational bottlenecks, we devise a mechanism to use augmented memory to scale NCA for large datasets and very deep networks. Our experiments deliver not only remarkable performance on ImageNet classification for such a simple non-parametric method, but most importantly a more generalizable feature representation for sub-category discovery and few-shot recognition.

Zhirong Wu, Yuanjun Xiong, Stella Yu et al.

Neural net classifiers trained on data with annotated class labels can also capture apparent visual similarity among categories without being directed to do so. We study whether this observation can be extended beyond the conventional domain of supervised learning: Can we learn a good feature representation that captures apparent similarity among instances, instead of classes, by merely asking the feature to be discriminative of individual instances? We formulate this intuition as a non-parametric classification problem at the instance-level, and use noise-contrastive estimation to tackle the computational challenges imposed by the large number of instance classes. Our experimental results demonstrate that, under unsupervised learning settings, our method surpasses the state-of-the-art on ImageNet classification by a large margin. Our method is also remarkable for consistently improving test performance with more training data and better network architectures. By fine-tuning the learned feature, we further obtain competitive results for semi-supervised learning and object detection tasks. Our non-parametric model is highly compact: With 128 features per image, our method requires only 600MB storage for a million images, enabling fast nearest neighbour retrieval at the run time.

Yue Zhao, Yuanjun Xiong, Limin Wang et al.

Detecting actions in untrimmed videos is an important yet challenging task. In this paper, we present the structured segment network (SSN), a novel framework which models the temporal structure of each action instance via a structured temporal pyramid. On top of the pyramid, we further introduce a decomposed discriminative model comprising two classifiers, respectively for classifying actions and determining completeness. This allows the framework to effectively distinguish positive proposals from background or incomplete ones, thus leading to both accurate recognition and localization. These components are integrated into a unified network that can be efficiently trained in an end-to-end fashion. Additionally, a simple yet effective temporal action proposal scheme, dubbed temporal actionness grouping (TAG) is devised to generate high quality action proposals. On two challenging benchmarks, THUMOS14 and ActivityNet, our method remarkably outperforms previous state-of-the-art methods, demonstrating superior accuracy and strong adaptivity in handling actions with various temporal structures.

Zhirong Wu, Dahua Lin, Xiaoou Tang

Markov Random Fields (MRFs), a formulation widely used in generative image modeling, have long been plagued by the lack of expressive power. This issue is primarily due to the fact that conventional MRFs formulations tend to use simplistic factors to capture local patterns. In this paper, we move beyond such limitations, and propose a novel MRF model that uses fully-connected neurons to express the complex interactions among pixels. Through theoretical analysis, we reveal an inherent connection between this model and recurrent neural networks, and thereon derive an approximated feed-forward network that couples multiple RNNs along opposite directions. This formulation combines the expressive power of deep neural networks and the cyclic dependency structure of MRF in a unified model, bringing the modeling capability to a new level. The feed-forward approximation also allows it to be efficiently learned from data. Experimental results on a variety of low-level vision tasks show notable improvement over state-of-the-arts.

Zhirong Wu, Dahua Lin, Xiaoou Tang

Binary representation is desirable for its memory efficiency, computation speed and robustness. In this paper, we propose adjustable bounded rectifiers to learn binary representations for deep neural networks. While hard constraining representations across layers to be binary makes training unreasonably difficult, we softly encourage activations to diverge from real values to binary by approximating step functions. Our final representation is completely binary. We test our approach on MNIST, CIFAR10, and ILSVRC2012 dataset, and systematically study the training dynamics of the binarization process. Our approach can binarize the last layer representation without loss of performance and binarize all the layers with reasonably small degradations. The memory space that it saves may allow more sophisticated models to be deployed, thus compensating the loss. To the best of our knowledge, this is the first work to report results on current deep network architectures using complete binary middle representations. Given the learned representations, we find that the firing or inhibition of a binary neuron is usually associated with a meaningful interpretation across different classes. This suggests that the semantic structure of a neural network may be manifested through a guided binarization process.

Zhirong Wu, Shuran Song, Aditya Khosla et al.

3D shape is a crucial but heavily underutilized cue in today's computer vision systems, mostly due to the lack of a good generic shape representation. With the recent availability of inexpensive 2.5D depth sensors (e.g. Microsoft Kinect), it is becoming increasingly important to have a powerful 3D shape representation in the loop. Apart from category recognition, recovering full 3D shapes from view-based 2.5D depth maps is also a critical part of visual understanding. To this end, we propose to represent a geometric 3D shape as a probability distribution of binary variables on a 3D voxel grid, using a Convolutional Deep Belief Network. Our model, 3D ShapeNets, learns the distribution of complex 3D shapes across different object categories and arbitrary poses from raw CAD data, and discovers hierarchical compositional part representations automatically. It naturally supports joint object recognition and shape completion from 2.5D depth maps, and it enables active object recognition through view planning. To train our 3D deep learning model, we construct ModelNet -- a large-scale 3D CAD model dataset. Extensive experiments show that our 3D deep representation enables significant performance improvement over the-state-of-the-arts in a variety of tasks.