Text-to-video generation poses significant challenges due to the inherent complexity of video data, which spans both temporal and spatial dimensions. It introduces additional redundancy, abrupt variations, and a domain gap between language and vision tokens while generation. Addressing these challenges requires an effective video tokenizer that can efficiently encode video data while preserving essential semantic and spatiotemporal information, serving as a critical bridge between text and vision. Inspired by the observation in VQ-VAE-2, we propose HiTVideo, a novel approach for text-to-video generation with hierarchical tokenizers. It utilizes a 3D causal VAE with a multi-layer discrete token framework, encoding video content into hierarchically structured codebooks. Higher layers capture semantic information with higher compression, while lower layers focus on fine-grained spatiotemporal details, striking a balance between compression efficiency and reconstruction quality. Our approach efficiently encodes longer video sequences (e.g., 8 seconds, 64 frames), reducing bits per pixel (bpp) by approximately 70% compared to previous tokenizers, while maintaining competitive reconstruction quality. We explore the trade-offs between compression and reconstruction, while emphasizing the advantages of high-compressed semantic tokens in text-to-video tasks. HiTVideo aims to address the potential limitations of existing video tokenizers in text-to-video generation tasks, striving for higher compression ratios, improved token quality, and simplify LLMs modeling under language guidance, offering a scalable and promising framework for advancing text to video generation.

Mingze Xu, Zijing Zhao, Qiming Peng, Houwen Peng, Han Hu, Zhanhui Kang, Yuxing Han. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent advancements in Large Language Models (LLMs) have demonstrated enhanced reasoning capabilities, evolving from Chainof-Thought (CoT) prompting to advanced, product-oriented solutions like OpenAI o1.During our re-implementation of this model, we noticed that in multimodal tasks requiring visual input (e.g., geometry problems), Multimodal LLMs (MLLMs) struggle to maintain focus on the visual information, in other words, MLLMs suffer from a gradual decline in attention to visual information as reasoning progresses, causing text-over-relied outputs.To investigate this, we ablate image inputs during long-chain reasoning.Concretely, we truncate the reasoning process midway, then re-complete the reasoning process with the input image removed.We observe only a ∼2% accuracy drop on MathVista's test-hard subset, revealing the model's textual outputs dominate the following reasoning process.Motivated by this, we propose Take-along Visual Conditioning (TVC), a strategy that shifts image input to critical reasoning stages and compresses redundant visual tokens via dynamic pruning.This methodology helps the model retain attention to the visual components throughout the reasoning.Our approach achieves state-of-the-art performance on average across five mathematical reasoning benchmarks (+3.4 points vs previous sota), demonstrating the effectiveness of TVC in enhancing multimodal reasoning systems.

Reasoning stands as a cornerstone of intelligence, enabling the synthesis of existing knowledge to solve complex problems. Despite remarkable progress, existing reasoning benchmarks often fail to rigorously evaluate the nuanced reasoning capabilities required for complex, real-world problemsolving, particularly in multi-disciplinary and multimodal contexts. In this paper, we introduce a graduate-level, multi-disciplinary, EnglishChinese benchmark, dubbed as Reasoning Bench (RBench), for assessing the reasoning capability of both language and multimodal models. RBench spans 1,094 questions across 108 subjects for language model evaluation and 665 questions across 83 subjects for multimodal model testing. These questions are meticulously curated to ensure rigorous difficulty calibration, subject balance, and cross-linguistic alignment, enabling the assessment to be an Olympiad-level multidisciplinary benchmark. We evaluate many models such as o1, GPT-4o, DeepSeek-R1, etc. Experimental results indicate that advanced models perform poorly on complex reasoning, especially multimodal reasoning. Even the top-performing model OpenAI o1 achieves only 53.2% accuracy on our multimodal evaluation. Data and code are made publicly available athttps://evalmodels.github.io/rbench/

The rapid advancement of native multi-modal models and omni-models, exemplified by GPT-4o, Gemini and o3 with their capability to process and generate content across modalities such as text and images, marks a significant milestone in the evolution of intelligence. Systematic evaluation of their multi-modal output capabilities in visual thinking process (a.k.a., multi-modal chain of thought, M-CoT) becomes critically important. However, existing benchmarks for evaluating multi-modal models primarily focus on assessing multi-modal inputs and text-only reasoning process while neglecting the importance of reasoning through multi-modal outputs. In this paper, we present a benchmark, dubbed as RBench-V, designed to assess models’ vision-indispensable reasoning. To conduct RBench-V, we carefully hand-pick 803 questions covering math, physics, counting and games. Unlike problems in previous benchmarks, which typically specify certain input modalities, RBench-V presents problems centered on multi-modal outputs, which require image manipulation, such as generating novel images and constructing auxiliary lines to support reasoning process. We evaluate numerous open- and closed-source models on RBench-V, including o3, Gemini 2.5 pro, Qwen2.5-VL, etc. Even the best-performing model, o3, achieves only 25.8% accuracy on RBench-V, far below the human score of 82.3%, which shows current models struggle to leverage multi-modal reasoning. Data and code are available at https://evalmodels.github.io/rbenchv.

High-quality data is crucial for the pre-training performance of large language models.Unfortunately, existing quality filtering methods rely on a known high-quality dataset as reference, which can introduce potential bias and compromise diversity.In this paper, we propose ScalingFilter, a novel approach that evaluates text quality based on the perplexity difference between two language models trained on the same data, thereby eliminating the influence of the reference dataset in the filtering process.An theoretical analysis shows that ScalingFilter is equivalent to an inverse utilization of scaling laws.Through training models with 1.3B parameters on the same data source processed by various quality filters, we find ScalingFilter can improve zero-shot performance of pre-trained models in downstream tasks.To assess the bias introduced by quality filtering, we introduce semantic diversity, a metric of utilizing text embedding models for semantic representations.Extensive experiments reveal that semantic diversity is a reliable indicator of dataset diversity, and ScalingFilter achieves an optimal balance between downstream performance and semantic diversity.1

In this paper, we present a new sequence-to-sequence learning framework for visual tracking, dubbed SeqTrack. It casts visual tracking as a sequence generation problem, which predicts object bounding boxes in an autoregressive fashion. This is different from prior Siamese trackers and transformer trackers, which rely on designing complicated head networks, such as classification and regression heads. SeqTrack only adopts a simple encoder-decoder transformer architecture. The encoder extracts visual features with a bidirectional transformer, while the decoder generates a sequence of bounding box values autoregressively with a causal transformer. The loss function is a plain cross-entropy. Such a sequence learning paradigm not only simplifies tracking framework, but also achieves competitive performance on benchmarks. For instance, SeqTrack gets 72.5% AUC on LaSOT, establishing a new state-of-the-art performance. Code and models are available at https://github.com/microsoft/VideoX.

Vision transformers have shown great success due to their high model capabilities. However, their remarkable performance is accompanied by heavy computation costs, which makes them unsuitable for real-time applications. In this paper, we propose a family of high-speed vision transformers named Efficient ViT. We find that the speed of existing transformer models is commonly bounded by memory inefficient operations, especially the tensor reshaping and element-wise functions in MHSA. Therefore, we design a new building block with a sandwich layout, i.e., using a single memory-bound MHSA between efficient FFN layers, which improves memory efficiency while enhancing channel communication. Moreover, we discover that the attention maps share high similarities across heads, leading to computational redundancy. To address this, we present a cascaded group attention module feeding attention heads with different splits of the full feature, which not only saves computation cost but also improves attention diversity. Comprehensive experiments demonstrate EfficientViT outperforms existing efficient models, striking a good trade-off between speed and accuracy. For instance, our EfficientViT-M5 surpasses MobileNetV3-Large by 1.9% in accuracy, while getting 40.4% and 45.2% higher throughput on Nvidia V100 GPU and Intel Xeon CPU, respectively. Compared to the recent efficient model MobileViT-XXS, EfficientViT-M2 achieves 1.8% superior accuracy, while running$5.8\times/3.7\times$faster on the GPU/CPU, and$7.4\times faster$when converted to ONNX format. Code and models are available at here.

This paper presents a method that effectively combines two prevalent visual recognition methods, i.e., image classification and contrastive language-image pre-training, dubbed iCLIP. Instead of naïve multi-task learning that use two separate heads for each task, we fuse the two tasks in a deep fashion that adapts the image classification to share the same formula and the same model weights with the language-image pre-training. To further bridge these two tasks, we propose to enhance the category names in image classification tasks using external knowledge, such as their descriptions in dictionaries. Extensive experiments show that the proposed method combines the advantages of two tasks well: the strong discrimination ability in image classification tasks due to the clean category labels, and the good zero-shot ability in CLIP tasks ascribed to the richer semantics in the text descriptions. In particular, it reaches 82.9% top-1 accuracy on IN-1K, and mean-while surpasses CLIP by 1.8%, with similar model size, on zero-shot recognition of Kornblith 12-dataset benchmark. The code and models are publicly available at https://github.com/weiyx16/iCLIP.

Transformer-based visual trackers have demonstrated significant progress owing to their superior modeling capabilities. However, existing trackers are hampered by low speed, limiting their applicability on devices with limited computational power. To alleviate this problem, we propose HiT, a new family of efficient tracking models that can run at high speed on different devices while retaining high performance. The central idea of HiT is the Bridge Module, which bridges the gap between modern lightweight transformers and the tracking framework. The Bridge Module incorporates the high-level information of deep features into the shallow large-resolution features. In this way, it produces better features for the tracking head. We also propose a novel dual-image position encoding technique that simultaneously encodes the position information of both the search region and template images. The HiT model achieves promising speed with competitive performance. For instance, it runs at 61 frames per second (fps) on the Nvidia Jetson AGX edge device. Furthermore, HiT attains 64.6% AUC on the LaSOT benchmark, surpassing all previous efficient trackers. Code and models are available at https://github.com/kangben258/HiT.

In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers’ behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by 50%, while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by 1.4 - 7.8× compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of 41.1% on ImageNet, surpassing the original CLIP ViT-B/16 by 3.5% while utilizing only 8.9% parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip.

In vision-language modeling, image token removal is an efficient augmentation technique to reduce the cost of encoding image features. The CLIP-style models, however, have been found to be negatively impacted by this technique. We hypothesize that removing a large portion of image tokens may inadvertently destroy the semantic information associated to a given text description, resulting in misaligned paired data in CLIP training. To address this issue, we propose an attentive token removal approach, which retains a small number of tokens that have a strong semantic correlation to the corresponding text description. The correlation scores are dynamically evaluated through an EMA-updated vision encoder. Our method, termed attentive mask CLIP, outperforms original CLIP and CLIP variant with random token removal while saving the training time. In addition, our approach also enables efficient multi-view contrastive learning. Experimentally, by training ViT-B on YFCC-15M dataset, our approach achieves 43.9% top-1 accuracy on ImageNet-1K zero-shot classification, 62.7/42.1 and 38.0/23.2 I2T/T2I retrieval accuracy on Flickr30K and MS COCO, outperforming SLIP by +1.1%, +5.5/+0.9, and +4.4/+1.3, respectively, while being 2.30× faster. An efficient version of our approach runs 1.16× faster than the plain CLIP model, while achieving significant gains of +5.3%, +11.3/+8.0, and +9.5/+4.9 on these benchmarks, respectively. Code will be release in https://github.com/microsoft/A-CLIP.

While language-guided image manipulation has made remarkable progress, the challenge of how to instruct the manipulation process faithfully reflecting human intentions persists. An accurate and comprehensive description of a manipulation task using natural language is laborious and sometimes even impossible, primarily due to the inherent uncertainty and ambiguity present in linguistic expressions. Is it feasible to accomplish image manipulation without resorting to external cross-modal language information? If this possibility exists, the inherent modality gap would be effortlessly eliminated. In this paper, we propose a novel manipulation methodology, dubbed ImageBrush, that learns visual instructions for more accurate image editing. Our key idea is to employ a pair of transformation images as visual instructions, which not only precisely captures human intention but also facilitates accessibility in real-world scenarios. Capturing visual instructions is particularly challenging because it involves extracting the underlying intentions solely from visual demonstrations and then applying this operation to a new image. To address this challenge, we formulate visual instruction learning as a diffusion-based inpainting problem, where the contextual information is fully exploited through an iterative process of generation. A visual prompting encoder is carefully devised to enhance the model's capacity in uncovering human intent behind the visual instructions. Extensive experiments show that our method generates engaging manipulation results conforming to the transformations entailed in demonstrations. Moreover, our model exhibits robust generalization capabilities on various downstream tasks such as pose transfer, image translation and video inpainting.

Differentiable ARchiTecture Search, i.e., DARTS, has drawn great attention in neural architecture search. It tries to find the optimal architecture in a shallow search network and then measures its performance in a deep evaluation network. The independent optimization of the search and evaluation networks, however, leaves a room for potential improvement by allowing interaction between the two networks. To address the problematic optimization issue, we propose new joint optimization objectives and a novel Cyclic Differentiable ARchiTecture Search framework, dubbed CDARTS. Considering the structure difference, CDARTS builds a cyclic feedback mechanism between the search and evaluation networks with introspective distillation. First, the search network generates an initial architecture for evaluation, and the weights of the evaluation network are optimized. Second, the architecture weights in the search network are further optimized by the label supervision in classification, as well as the regularization from the evaluation network through feature distillation. Repeating the above cycle results in a joint optimization of the search and evaluation networks and thus enables the evolution of the architecture to fit the final evaluation network. The experiments and analysis on CIFAR, ImageNet and NATS-Bench [95] demonstrate the effectiveness of the proposed approach over the state-of-the-art ones. Specifically, in the DARTS search space, we achieve 97.52% top-1 accuracy on CIFAR10 and 76.3% top-1 accuracy on ImageNet. In the chain-structured search space, we achieve 78.2% top-1 accuracy on ImageNet, which is 1.1% higher than EfficientNet-B0. Our code and models are publicly available at https://github.com/microsoft/Cream.

Vision Transformer (ViT) models have recently drawn much attention in computer vision due to their high model capability. However, ViT models suffer from huge number of parameters, restricting their applicability on devices with limited memory. To alleviate this problem, we propose MiniViT, a new compression framework, which achieves parameter reduction in vision transformers while retaining the same performance. The central idea of MiniViT is to multiplex the weights of consecutive transformer blocks. More specifically, we make the weights shared across layers, while imposing a transformation on the weights to increase diversity. Weight distillation over self-attention is also applied to transfer knowledge from large-scale ViT models to weight-multiplexed compact models. Comprehensive experiments demonstrate the efficacy of MiniViT, showing that it can reduce the size of the pre-trained Swin-B transformer by 48%, while achieving an increase of 1.0% in Top-1 accuracy on ImageNet. Moreover, using a single-layer of parameters, MiniViT is able to compress DeiT-B by 9.7 times from 86M to 9M parameters, without seriously compromising the performance. Finally, we verify the transferability of MiniViT by reporting its performance on downstream benchmarks. Code and models are available at here.

PointNet++ is one of the most influential neural architectures for point cloud understanding. Although the accuracy of PointNet++ has been largely surpassed by recent networks such as PointMLP and Point Transformer, we find that a large portion of the performance gain is due to improved training strategies, i.e. data augmentation and optimization techniques, and increased model sizes rather than architectural innovations. Thus, the full potential of PointNet++ has yet to be explored. In this work, we revisit the classical PointNet++ through a systematic study of model training and scaling strategies, and offer two major contributions. First, we propose a set of improved training strategies that significantly improve PointNet++ performance. For example, we show that, without any change in architecture, the overall accuracy (OA) of PointNet++ on ScanObjectNN object classification can be raised from 77.9% to 86.1%, even outperforming state-of-the-art PointMLP. Second, we introduce an inverted residual bottleneck design and separable MLPs into PointNet++ to enable efficient and effective model scaling and propose PointNeXt, the next version of PointNets. PointNeXt can be flexibly scaled up and outperforms state-of-the-art methods on both 3D classification and segmentation tasks. For classification, PointNeXt reaches an overall accuracy of 87.7 on ScanObjectNN, surpassing PointMLP by 2.3%, while being 10x faster in inference. For semantic segmentation, PointNeXt establishes a new state-of-the-art performance with 74.9% mean IoU on S3DIS (6-fold cross-validation), being superior to the recent Point Transformer. The code and models are available at https://github.com/guochengqian/pointnext.

We address the problem of retrieving a specific moment from an untrimmed video by natural language. It is a challenging problem because a target moment may take place in the context of other temporal moments in the untrimmed video. Existing methods cannot tackle this challenge well since they do not fully consider the temporal contexts between temporal moments. In this paper, we model the temporal context between video moments by a set of predefined two-dimensional maps under different temporal scales. For each map, one dimension indicates the starting time of a moment and the other indicates the duration. These 2D temporal maps can cover diverse video moments with different lengths, while representing their adjacent contexts at different temporal scales. Based on the 2D temporal maps, we propose a Multi-Scale Temporal Adjacency Network (MS-2D-TAN), a single-shot framework for moment localization. It is capable of encoding the adjacent temporal contexts at each scale, while learning discriminative features for matching video moments with referring expressions. We evaluate the proposed MS-2D-TAN on three challenging benchmarks, i.e., Charades-STA, ActivityNet Captions, and TACoS, where our MS-2D-TAN outperforms the state of the art.

Object tracking has achieved significant progress over the past few years. However, state-of-the-art trackers become increasingly heavy and expensive, which limits their deployments in resource-constrained applications. In this work, we present LightTrack, which uses neural architecture search (NAS) to design more lightweight and efficient object trackers. Comprehensive experiments show that our LightTrack is effective. It can find trackers that achieve superior performance compared to handcrafted SOTA trackers, such as SiamRPN++ [30] and Ocean [56], while using much fewer model Flops and parameters. Moreover, when deployed on resource-constrained mobile chipsets, the discovered trackers run much faster. For example, on Snapdragon 845 Adreno GPU, LightTrack runs 12× faster than Ocean, while using 13× fewer parameters and 38× fewer Flops. Such improvements might narrow the gap between academic models and industrial deployments in object tracking task. LightTrack is released at here.

In this paper, we present a new tracking architecture with an encoder-decoder transformer as the key component. The encoder models the global spatio-temporal feature dependencies between target objects and search regions, while the decoder learns a query embedding to predict the spatial positions of the target objects. Our method casts object tracking as a direct bounding box prediction problem, without using any proposals or predefined anchors. With the encoder-decoder transformer, the prediction of objects just uses a simple fully-convolutional network, which estimates the corners of objects directly. The whole method is end-to-end, does not need any postprocessing steps such as cosine window and bounding box smoothing, thus largely simplifying existing tracking pipelines. The proposed tracker achieves state-of-the-art performance on multiple challenging short-term and long-term benchmarks, while running at real-time speed, being 6× faster than Siam R-CNN [54]. Code and models are open-sourced at https://github.com/researchmm/Stark.

Recently, pure transformer-based models have shown great potentials for vision tasks such as image classification and detection. However, the design of transformer networks is challenging. It has been observed that the depth, embedding dimension, and number of heads can largely affect the performance of vision transformers. Previous models configure these dimensions based upon manual crafting. In this work, we propose a new one-shot architecture search framework, namely AutoFormer, dedicated to vision transformer search. AutoFormer entangles the weights of different blocks in the same layers during supernet training. Benefiting from the strategy, the trained supernet allows thousands of subnets to be very well-trained. Specifically, the performance of these subnets with weights inherited from the supernet is comparable to those retrained from scratch. Besides, the searched models, which we refer to AutoFormers, surpass the recent state-of-the-arts such as ViT and DeiT. In particular, AutoFormer-tiny/small/base achieve 74.7%/81.7%/82.4% top-1 accuracy on ImageNet with 5.7M/22.9M/53.7M parameters, respectively. Lastly, we verify the transferability of AutoFormer by providing the performance on downstream benchmarks and distillation experiments. Code and models are available at https://github.com/microsoft/Cream.

Relative position encoding (RPE) is important for transformer to capture sequence ordering of input tokens. General efficacy has been proven in natural language processing. However, in computer vision, its efficacy is not well studied and even remains controversial, e.g., whether relative position encoding can work equally well as absolute position? In order to clarify this, we first review existing relative position encoding methods and analyze their pros and cons when applied in vision transformers. We then propose new relative position encoding methods dedicated to 2D images, called image RPE (iRPE). Our methods consider directional relative distance modeling as well as the interactions between queries and relative position embeddings in self-attention mechanism. The proposed iRPE methods are simple and lightweight. They can be easily plugged into transformer blocks. Experiments demonstrate that solely due to the proposed encoding methods, DeiT [21] and DETR [1] obtain up to 1.5% (top-1 Acc) and 1.3% (mAP) stable improvements over their original versions on ImageNet and COCO respectively, without tuning any extra hyperparameters such as learning rate and weight decay. Our ablation and analysis also yield interesting findings, some of which run counter to previous understanding. Code and models are open-sourced at https://github.com/microsoft/Cream/tree/main/iRPE.

In this paper, we propose to learn an Unsupervised Single Object Tracker (USOT) from scratch. We identify that three major challenges, i.e., moving object discovery, rich temporal variation exploitation, and online update, are the central causes of the performance bottleneck of existing unsupervised trackers. To narrow the gap between unsupervised trackers and supervised counterparts, we propose an effective unsupervised learning approach composed of three stages. First, we sample sequentially moving objects with unsupervised optical flow and dynamic programming, instead of random cropping. Second, we train a naive Siamese tracker from scratch using single-frame pairs. Third, we continue training the tracker with a novel cycle memory learning scheme, which is conducted in longer temporal spans and also enables our tracker to update online. Extensive experiments show that the proposed USOT learned from unlabeled videos performs well over the state-of-the-art unsupervised trackers by large margins, and on par with recent supervised deep trackers. Code is available at https://github.com/VISION-SJTU/USOT.

Vision Transformer has shown great visual representation power in substantial vision tasks such as recognition and detection, and thus been attracting fast-growing efforts on manually designing more effective architectures. In this paper, we propose to use neural architecture search to automate this process, by searching not only the architecture but also the search space. The central idea is to gradually evolve different search dimensions guided by their E-T Error computed using a weight-sharing supernet. Moreover, we provide design guidelines of general vision transformers with extensive analysis according to the space searching process, which could promote the understanding of vision transformer. Remarkably, the searched models, named S3 (short for Searching the Search Space), from the searched space achieve superior performance to recently proposed models, such as Swin, DeiT and ViT, when evaluated on ImageNet. The effectiveness of S3 is also illustrated on object detection, semantic segmentation and visual question answering, demonstrating its generality to downstream vision and vision-language tasks. Code and models will be available at https://github.com/microsoft/Cream.

Vision-Language Pre-training (VLP) aims to learn multi-modal representations from image-text pairs and serves for downstream vision-language tasks in a fine-tuning fashion. The dominant VLP models adopt a CNN-Transformer architecture, which embeds images with a CNN, and then aligns images and text with a Transformer. Visual relationship between visual contents plays an important role in image understanding and is the basic for inter-modal alignment learning. However, CNNs have limitations in visual relation learning due to local receptive field's weakness in modeling long-range dependencies. Thus the two objectives of learning visual relation and inter-modal alignment are encapsulated in the same Transformer network. Such design might restrict the inter-modal alignment learning in the Transformer by ignoring the specialized characteristic of each objective. To tackle this, we propose a fully Transformer visual embedding for VLP to better learn visual relation and further promote inter-modal alignment. Specifically, we propose a metric named Inter-Modality Flow (IMF) to measure the interaction between vision and language modalities (i.e., inter-modality). We also design a novel masking optimization mechanism named Masked Feature Regression (MFR) in Transformer to further promote the inter-modality learning. To the best of our knowledge, this is the first study to explore the benefit of Transformer for visual feature learning in VLP. We verify our method on a wide range of vision-language tasks, including Visual Question Answering (VQA), Visual Entailment and Visual Reasoning. Our approach not only outperforms the state-of-the-art VLP performance, but also shows benefits on the IMF metric.

Pixelwise single object tracking is challenging due to the competition of running speeds and segmentation accuracy. Current state-of-the-art real-time approaches seamlessly connect tracking and segmentation by sharing computation of the backbone network, e.g., SiamMask and D3S fork a light branch from the tracking model to predict segmentation mask. Although efficient, directly reusing features from tracking networks may harm the segmentation accuracy, since background clutter in the backbone feature tends to introduce false positives in segmentation. To mitigate this problem, we propose a unified tracking-retrieval-segmentation framework consisting of an attention retrieval network (ARN) and an iterative feedback network (IFN). Instead of segmenting the target inside the bounding box, the proposed framework performs soft spatial constraints on backbone features to obtain an accurate global segmentation map. Concretely, in ARN, a look-up-table (LUT) is first built by sufficiently using the information of the first frame. By retrieving it, a target-aware attention map is generated to suppress the negative influence of background clutter. To ulteriorly refine the contour of the segmentation, IFN iteratively enhances the features at different resolutions by taking the predicted mask as feedback guidance. Our framework sets a new state of the art on the recent pixelwise tracking benchmark VOT2020 and runs at 40 fps. Notably, the proposed model surpasses SiamMask by 11.7/4.2/5.5 points on VOT2020, DAVIS2016, and DAVIS2017, respectively. Code is available at https://github.com/JudasDie/SOTS.

We address the problem of retrieving a specific moment from an untrimmed video by a query sentence. This is a challenging problem because a target moment may take place in relations to other temporal moments in the untrimmed video. Existing methods cannot tackle this challenge well since they consider temporal moments individually and neglect the temporal dependencies. In this paper, we model the temporal relations between video moments by a two-dimensional map, where one dimension indicates the starting time of a moment and the other indicates the end time. This 2D temporal map can cover diverse video moments with different lengths, while representing their adjacent relations. Based on the 2D map, we propose a Temporal Adjacent Network (2D-TAN), a single-shot framework for moment localization. It is capable of encoding the adjacent temporal relation, while learning discriminative features for matching video moments with referring expressions. We evaluate the proposed 2D-TAN on three challenging benchmarks, i.e., Charades-STA, ActivityNet Captions, and TACoS, where our 2D-TAN outperforms the state-of-the-art.

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 ~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 https://github.com/ microsoft/transductive-vos.pytorch.

One-shot weight sharing methods have recently drawn great attention in neural architecture search due to high efficiency and competitive performance. However, weight sharing across models has an inherent deficiency, i.e., insufficient training of subnetworks in the hypernetwork. To alleviate this problem, we present a simple yet effective architecture distillation method. The central idea is that subnetworks can learn collaboratively and teach each other throughout the training process, aiming to boost the convergence of individual models. We introduce the concept of prioritized path, which refers to the architecture candidates exhibiting superior performance during training. Distilling knowledge from the prioritized paths is able to boost the training of subnetworks. Since the prioritized paths are changed on the fly depending on their performance and complexity, the final obtained paths are the cream of the crop. We directly select the most promising one from the prioritized paths as the final architecture, without using other complex search methods, such as reinforcement learning or evolution algorithms. The experiments on ImageNet verify such path distillation method can improve the convergence ratio and performance of the hypernetwork, as well as boosting the training of subnetworks. The discovered architectures achieve superior performance compared to the recent MobileNetV3 and EfficientNet families under aligned settings. Moreover, the experiments on object detection and more challenging search space show the generality and robustness of the proposed method. Code and models are available at \url{https://github.com/neurips-20/cream.git}.

Most of the current action localization methods follow an anchor-based pipeline: depicting action instances by pre-defined anchors, learning to select the anchors closest to the ground truth, and predicting the confidence of anchors with refinements. Pre-defined anchors set prior about the location and duration for action instances, which facilitates the localization for common action instances but limits the flexibility for tackling action instances with drastic varieties, especially for extremely short or extremely long ones. To address this problem, this paper proposes a novel anchor-free action localization module that assists action localization by temporal points. Specifically, this module represents an action instance as a point with its distances to the starting boundary and ending boundary, alleviating the pre-defined anchor restrictions in terms of action localization and duration. The proposed anchor-free module is capable of predicting the action instances whose duration is either extremely short or extremely long. By combining the proposed anchor-free module with a conventional anchor-based module, we propose a novel action localization framework, called A2Net. The cooperation between anchor-free and anchor-based modules achieves superior performance to the state-of-the-art on THUMOS14 (45.5% vs. 42.8%). Furthermore, comprehensive experiments demonstrate the complementarity between the anchor-free and the anchor-based module, making A2Net simple but effective.

Siamese networks have drawn great attention in visual tracking because of their balanced accuracy and speed. However, the backbone networks used in Siamese trackers are relatively shallow, such as AlexNet, which does not fully take advantage of the capability of modern deep neural networks. In this paper, we investigate how to leverage deeper and wider convolutional neural networks to enhance tracking robustness and accuracy. We observe that direct replacement of backbones with existing powerful architectures, such as ResNet and Inception, does not bring improvements. The main reasons are that 1) large increases in the receptive field of neurons lead to reduced feature discriminability and localization precision; and 2) the network padding for convolutions induces a positional bias in learning. To address these issues, we propose new residual modules to eliminate the negative impact of padding, and further design new architectures using these modules with controlled receptive field size and network stride. The designed architectures are lightweight and guarantee real-time tracking speed when applied to SiamFC and SiamRPN. Experiments show that solely due to the proposed network architectures, our SiamFC+ and SiamRPN+ obtain up to 9.8%/6.3% (AUC), 23.3%/8.8% (EAO) and 24.4%/25.0% (EAO) relative improvements over the original versions on the OTB-15, VOT-16 and VOT-17 datasets, respectively.

Recent years have witnessed an unprecedented growing of sport videos, as different types of sports activities can be widely-observed (i.e., from professional athletics to personal fitness). Existing approaches by computer vision have predominantly focused on creating experiences of content browsing and searching by video tagging and summarization. These techniques have already enabled a wide-range of applications for sports enthusiasts, such as text-based video search, highlight generation, and so on. In this paper, we take one step further to create an AI coach system to provide personalized athletic training experiences. Especially for sports activities which the training quality largely depends on the correctness of human poses in a video sequence. As sports videos often involve grand challenges of fast movement (e.g., skiing, skating) and complex actions (e.g., gymnastics), we propose to design the system with several distinct features: (1) trajectory extraction for a single human instance by leveraging deep visual tracking, (2) human pose estimation by proposing a novel human joints relation model in spatial and temporal domains, (3) pose correction by abnormal detection and exemplar-based visual suggestions. We have collected sports training videos from 30 sports enthusiasts, namely Freestyle Skiing Aerials dataset (63 clips). We show that the proposed system can lead to a remarkably better user training experience by extensive user studies.

Accurate pose analysis in sport videos is beneficial to users to improve skills. In this paper, we propose an AI coach system to provide personalized athletic training experiences for posture-wise sports activities, in which the training quality largely depends on the correctness of human poses in a video sequence. we propose to design the system with several distinct features: (1) trajectory extraction for a single human instance by leveraging deep visual tracking, (2) human pose estimation by proposing a novel human joints relation model in spatial and temporal domains,(3) pose correction by abnormal detection, performance rating and exemplar-based visual suggestions. We build an online service of this AI coach system for sports enthusiasts and collect extensive feedbacks. Comparisons with some latest popular sport apps demonstrate the effectiveness of this AI coach system to improve skills for users.

In multi-instance learning (MIL), the relations among instances in a bag convey important contextual information in many applications. Previous studies on MIL either ignore such relations or simply model them with a fixed graph structure so that the overall performance inevitably degrades in complex environments. To address this problem, this paper proposes a novel multi-view multi-instance learning algorithm (MIL) that combines multiple context structures in a bag into a unified framework. The novel aspects are: (i) we propose a sparse -graph model that can generate different graphs with different parameters to represent various context relations in a bag, (ii) we propose a multi-view joint sparse representation that integrates these graphs into a unified framework for bag classification, and (iii) we propose a multi-view dictionary learning algorithm to obtain a multi-view graph dictionary that considers cues from all views simultaneously to improve the discrimination of the MIL. Experiments and analyses in many practical applications prove the effectiveness of the M IL.

Low-rank recovery models have shown potential for salient object detection, where a matrix is decomposed into a low-rank matrix representing image background and a sparse matrix identifying salient objects. Two deficiencies, however, still exist. First, previous work typically assumes the elements in the sparse matrix are mutually independent, ignoring the spatial and pattern relations of image regions. Second, when the low-rank and sparse matrices are relatively coherent, e.g., when there are similarities between the salient objects and background or when the background is complicated, it is difficult for previous models to disentangle them. To address these problems, we propose a novel structured matrix decomposition model with two structural regularizations: (1) a tree-structured sparsity-inducing regularization that captures the image structure and enforces patches from the same object to have similar saliency values, and (2) a Laplacian regularization that enlarges the gaps between salient objects and the background in feature space. Furthermore, high-level priors are integrated to guide the matrix decomposition and boost the detection. We evaluate our model for salient object detection on five challenging datasets including single object, multiple objects and complex scene images, and show competitive results as compared with 24 state-of-the-art methods in terms of seven performance metrics.

The images we encounter throughout our lives make different impressions on us: Some are remembered at first glance, while others are forgotten. This phenomenon is caused by the intrinsic memorability of images revealed by recent studies [5,6]. In this paper, we address the issue of automatically estimating the memorability of images by proposing a novel multi-view adaptive regression (MAR) model. The MAR model provides an effective mapping of visual features to memorability scores by taking advantage of robust feature selection and multiple feature integration. It consists of three major components: an adaptive loss function, an adaptive regularization and a multi-view modeling strategy. Moreover, we design an alternating direction method (ADM) optimization algorithm to solve the proposed objective function. Experimental results on the MIT benchmark dataset show the superiority of the proposed model compared with existing image memorability prediction methods.

Salient object detection provides an alternative solution to various image semantic understanding tasks such as object recognition, adaptive compression and image retrieval. Recently, low-rank matrix recovery (LR) theory has been introduced into saliency detection, and achieves impressed results. However, the existing LR-based models neglect the underlying structure of images, and inevitably degrade the associated performance. In this paper, we propose a Low-rank and Structured sparse Matrix Decomposition (LSMD) model for salient object detection. In the model, a tree-structured sparsity-inducing norm regularization is firstly introduced to provide a hierarchical description of the image structure to ensure the completeness of the extracted salient object. The similarity of saliency values within the salient object is then guaranteed by the $\ell _\infty$-norm. Finally, high-level priors are integrated to guide the matrix decomposition and enhance the saliency detection. Experimental results on the largest public benchmark database show that our model outperforms existing LR-based approaches and other state-of-the-art methods, which verifies the effectiveness and robustness of the structure cues in our model.

Computational color constancy is a very important topic in computer vision and has attracted many researchers' attention. Recently, lots of research has shown the effects of using high level visual content cues for improving illumination estimation. However, nearly all the existing methods are essentially combinational strategies in which image's content analysis is only used to guide the combination or selection from a variety of individual illumination estimation methods. In this paper, we propose a novel bilayer sparse coding model for illumination estimation that considers image similarity in terms of both low level color distribution and high level image scene content simultaneously. For the purpose, the image's scene content information is integrated with its color distribution to obtain optimal illumination estimation model. The experimental results on real-world image sets show that our algorithm is superior to some prevailing illumination estimation methods, even better than some combinational methods.