Lei Zhang

Yuxiang Wei, Yabo Zhang, Zhilong Ji et al. · microsoft-research

In addition to the unprecedented ability in imaginary creation, large text-to-image models are expected to take customized concepts in image generation. Existing works generally learn such concepts in an optimization-based manner, yet bringing excessive computation or memory burden. In this paper, we instead propose a learning-based encoder, which consists of a global and a local mapping networks for fast and accurate customized text-to-image generation. In specific, the global mapping network projects the hierarchical features of a given image into multiple new words in the textual word embedding space, i.e., one primary word for well-editable concept and other auxiliary words to exclude irrelevant disturbances (e.g., background). In the meantime, a local mapping network injects the encoded patch features into cross attention layers to provide omitted details, without sacrificing the editability of primary concepts. We compare our method with existing optimization-based approaches on a variety of user-defined concepts, and demonstrate that our method enables high-fidelity inversion and more robust editability with a significantly faster encoding process. Our code is publicly available at https://github.com/csyxwei/ELITE.

Yabin Zhang, Minghan Li, Ruihuang Li et al. · microsoft-research, stanford

Arbitrary style transfer (AST) and domain generalization (DG) are important yet challenging visual learning tasks, which can be cast as a feature distribution matching problem. With the assumption of Gaussian feature distribution, conventional feature distribution matching methods usually match the mean and standard deviation of features. However, the feature distributions of real-world data are usually much more complicated than Gaussian, which cannot be accurately matched by using only the first-order and second-order statistics, while it is computationally prohibitive to use high-order statistics for distribution matching. In this work, we, for the first time to our best knowledge, propose to perform Exact Feature Distribution Matching (EFDM) by exactly matching the empirical Cumulative Distribution Functions (eCDFs) of image features, which could be implemented by applying the Exact Histogram Matching (EHM) in the image feature space. Particularly, a fast EHM algorithm, named Sort-Matching, is employed to perform EFDM in a plug-and-play manner with minimal cost. The effectiveness of our proposed EFDM method is verified on a variety of AST and DG tasks, demonstrating new state-of-the-art results. Codes are available at https://github.com/YBZh/EFDM.

Hao Zhang, Feng Li, Shilong Liu et al. · tsinghua

We present DINO (\textbf{D}ETR with \textbf{I}mproved de\textbf{N}oising anch\textbf{O}r boxes), a state-of-the-art end-to-end object detector. % in this paper. DINO improves over previous DETR-like models in performance and efficiency by using a contrastive way for denoising training, a mixed query selection method for anchor initialization, and a look forward twice scheme for box prediction. DINO achieves $49.4$AP in $12$ epochs and $51.3$AP in $24$ epochs on COCO with a ResNet-50 backbone and multi-scale features, yielding a significant improvement of $\textbf{+6.0}$\textbf{AP} and $\textbf{+2.7}$\textbf{AP}, respectively, compared to DN-DETR, the previous best DETR-like model. DINO scales well in both model size and data size. Without bells and whistles, after pre-training on the Objects365 dataset with a SwinL backbone, DINO obtains the best results on both COCO \texttt{val2017} ($\textbf{63.2}$\textbf{AP}) and \texttt{test-dev} (\textbf{$\textbf{63.3}$AP}). Compared to other models on the leaderboard, DINO significantly reduces its model size and pre-training data size while achieving better results. Our code will be available at \url{https://github.com/IDEACVR/DINO}.

Feng Li, Hao Zhang, Shilong Liu et al. · tsinghua

We present in this paper a novel denoising training method to speedup DETR (DEtection TRansformer) training and offer a deepened understanding of the slow convergence issue of DETR-like methods. We show that the slow convergence results from the instability of bipartite graph matching which causes inconsistent optimization goals in early training stages. To address this issue, except for the Hungarian loss, our method additionally feeds ground-truth bounding boxes with noises into Transformer decoder and trains the model to reconstruct the original boxes, which effectively reduces the bipartite graph matching difficulty and leads to a faster convergence. Our method is universal and can be easily plugged into any DETR-like methods by adding dozens of lines of code to achieve a remarkable improvement. As a result, our DN-DETR results in a remarkable improvement ($+1.9$AP) under the same setting and achieves the best result (AP $43.4$ and $48.6$ with $12$ and $50$ epochs of training respectively) among DETR-like methods with ResNet-$50$ backbone. Compared with the baseline under the same setting, DN-DETR achieves comparable performance with $50\%$ training epochs. Code is available at \url{https://github.com/FengLi-ust/DN-DETR}.

Xinyu Huang, Youcai Zhang, Jinyu Ma et al. · microsoft-research

This paper presents Tag2Text, a vision language pre-training (VLP) framework, which introduces image tagging into vision-language models to guide the learning of visual-linguistic features. In contrast to prior works which utilize object tags either manually labeled or automatically detected with an off-the-shelf detector with limited performance, our approach explicitly learns an image tagger using tags parsed from image-paired text and thus provides a strong semantic guidance to vision-language models. In this way, Tag2Text can utilize large-scale annotation-free image tags in accordance with image-text pairs, and provides more diverse tag categories beyond objects. As a result, Tag2Text demonstrates the ability of a foundational image tagging model, with superior zero-shot performance even comparable to fully supervised models. Moreover, by leveraging the tagging guidance, Tag2Text effectively enhances the performance of vision-language models on both generation-based and alignment-based tasks. Across a wide range of downstream benchmarks, Tag2Text achieves state-of-the-art results with similar model sizes and data scales, demonstrating the efficacy of the proposed tagging guidance. Code, demo and pre-trained models are available at https://github.com/xinyu1205/recognize-anything.

Feng Li, Hao Zhang, Huaizhe xu et al. · tsinghua

In this paper we present Mask DINO, a unified object detection and segmentation framework. Mask DINO extends DINO (DETR with Improved Denoising Anchor Boxes) by adding a mask prediction branch which supports all image segmentation tasks (instance, panoptic, and semantic). It makes use of the query embeddings from DINO to dot-product a high-resolution pixel embedding map to predict a set of binary masks. Some key components in DINO are extended for segmentation through a shared architecture and training process. Mask DINO is simple, efficient, and scalable, and it can benefit from joint large-scale detection and segmentation datasets. Our experiments show that Mask DINO significantly outperforms all existing specialized segmentation methods, both on a ResNet-50 backbone and a pre-trained model with SwinL backbone. Notably, Mask DINO establishes the best results to date on instance segmentation (54.5 AP on COCO), panoptic segmentation (59.4 PQ on COCO), and semantic segmentation (60.8 mIoU on ADE20K) among models under one billion parameters. Code is available at \url{https://github.com/IDEACVR/MaskDINO}.

Shilong Liu, Zhaoyang Zeng, Tianhe Ren et al. · tsinghua

In this paper, we present an open-set object detector, called Grounding DINO, by marrying Transformer-based detector DINO with grounded pre-training, which can detect arbitrary objects with human inputs such as category names or referring expressions. The key solution of open-set object detection is introducing language to a closed-set detector for open-set concept generalization. To effectively fuse language and vision modalities, we conceptually divide a closed-set detector into three phases and propose a tight fusion solution, which includes a feature enhancer, a language-guided query selection, and a cross-modality decoder for cross-modality fusion. While previous works mainly evaluate open-set object detection on novel categories, we propose to also perform evaluations on referring expression comprehension for objects specified with attributes. Grounding DINO performs remarkably well on all three settings, including benchmarks on COCO, LVIS, ODinW, and RefCOCO/+/g. Grounding DINO achieves a $52.5$ AP on the COCO detection zero-shot transfer benchmark, i.e., without any training data from COCO. It sets a new record on the ODinW zero-shot benchmark with a mean $26.1$ AP. Code will be available at \url{https://github.com/IDEA-Research/GroundingDINO}.

Yichen Yuan, Yifan Wang, Lijun Wang et al. · stanford

Recent leading zero-shot video object segmentation (ZVOS) works devote to integrating appearance and motion information by elaborately designing feature fusion modules and identically applying them in multiple feature stages. Our preliminary experiments show that with the strong long-range dependency modeling capacity of Transformer, simply concatenating the two modality features and feeding them to vanilla Transformers for feature fusion can distinctly benefit the performance but at a cost of heavy computation. Through further empirical analysis, we find that attention dependencies learned in Transformer in different stages exhibit completely different properties: global query-independent dependency in the low-level stages and semantic-specific dependency in the high-level stages. Motivated by the observations, we propose two Transformer variants: i) Context-Sharing Transformer (CST) that learns the global-shared contextual information within image frames with a lightweight computation. ii) Semantic Gathering-Scattering Transformer (SGST) that models the semantic correlation separately for the foreground and background and reduces the computation cost with a soft token merging mechanism. We apply CST and SGST for low-level and high-level feature fusions, respectively, formulating a level-isomerous Transformer framework for ZVOS task. Compared with the baseline that uses vanilla Transformers for multi-stage fusion, ours significantly increase the speed by 13 times and achieves new state-of-the-art ZVOS performance. Code is available at https://github.com/DLUT-yyc/Isomer.

Minghan Li, Shuai Li, Wangmeng Xiang et al. · microsoft-research

While impressive progress has been achieved, video instance segmentation (VIS) methods with per-clip input often fail on challenging videos with occluded objects and crowded scenes. This is mainly because instance queries in these methods cannot encode well the discriminative embeddings of instances, making the query-based segmenter difficult to distinguish those `hard' instances. To address these issues, we propose to mine discriminative query embeddings (MDQE) to segment occluded instances on challenging videos. First, we initialize the positional embeddings and content features of object queries by considering their spatial contextual information and the inter-frame object motion. Second, we propose an inter-instance mask repulsion loss to distance each instance from its nearby non-target instances. The proposed MDQE is the first VIS method with per-clip input that achieves state-of-the-art results on challenging videos and competitive performance on simple videos. In specific, MDQE with ResNet50 achieves 33.0\% and 44.5\% mask AP on OVIS and YouTube-VIS 2021, respectively. Code of MDQE can be found at \url{https://github.com/MinghanLi/MDQE_CVPR2023}.

Guanyi Qin, Jie Liang, Bingbing Zhang et al. · baidu

In this paper, we present an overview of the NTIRE 2026 challenge on the 3rd Restore Any Image Model in the Wild, specifically focusing on Track 1: Professional Image Quality Assessment. Conventional Image Quality Assessment (IQA) typically relies on scalar scores. By compressing complex visual characteristics into a single number, these methods fundamentally struggle to distinguish subtle differences among uniformly high-quality images. Furthermore, they fail to articulate why one image is superior, lacking the reasoning capabilities required to provide guidance for vision tasks. To bridge this gap, recent advancements in Multimodal Large Language Models (MLLMs) offer a promising paradigm. Inspired by this potential, our challenge establishes a novel benchmark exploring the ability of MLLMs to mimic human expert cognition in evaluating high-quality image pairs. Participants were tasked with overcoming critical bottlenecks in professional scenarios, centering on two primary objectives: (1) Comparative Quality Selection: reliably identifying the visually superior image within a high-quality pair; and (2) Interpretative Reasoning: generating grounded, expert-level explanations that detail the rationale behind the selection. In total, the challenge attracted nearly 200 registrations and over 2,500 submissions. The top-performing methods significantly advanced the state of the art in professional IQA. The challenge dataset is available at https://github.com/narthchin/RAIM-PIQA, and the official homepage is accessible at https://www.codabench.org/competitions/12789/.

Ruihuang Li, Chenhang He, Shuai Li et al. · stanford

The representative instance segmentation methods mostly segment different object instances with a mask of the fixed resolution, e.g., 28*28 grid. However, a low-resolution mask loses rich details, while a high-resolution mask incurs quadratic computation overhead. It is a challenging task to predict the optimal binary mask for each instance. In this paper, we propose to dynamically select suitable masks for different object proposals. First, a dual-level Feature Pyramid Network (FPN) with adaptive feature aggregation is developed to gradually increase the mask grid resolution, ensuring high-quality segmentation of objects. Specifically, an efficient region-level top-down path (r-FPN) is introduced to incorporate complementary contextual and detailed information from different stages of image-level FPN (i-FPN). Then, to alleviate the increase of computation and memory costs caused by using large masks, we develop a Mask Switch Module (MSM) with negligible computational cost to select the most suitable mask resolution for each instance, achieving high efficiency while maintaining high segmentation accuracy. Without bells and whistles, the proposed method, namely DynaMask, brings consistent and noticeable performance improvements over other state-of-the-arts at a moderate computation overhead. The source code: https://github.com/lslrh/DynaMask.

Yabin Zhang, Jiehong Lin, Chenhang He et al. · stanford

Masked auto-encoding is a popular and effective self-supervised learning approach to point cloud learning. However, most of the existing methods reconstruct only the masked points and overlook the local geometry information, which is also important to understand the point cloud data. In this work, we make the first attempt, to the best of our knowledge, to consider the local geometry information explicitly into the masked auto-encoding, and propose a novel Masked Surfel Prediction (MaskSurf) method. Specifically, given the input point cloud masked at a high ratio, we learn a transformer-based encoder-decoder network to estimate the underlying masked surfels by simultaneously predicting the surfel positions (i.e., points) and per-surfel orientations (i.e., normals). The predictions of points and normals are supervised by the Chamfer Distance and a newly introduced Position-Indexed Normal Distance in a set-to-set manner. Our MaskSurf is validated on six downstream tasks under three fine-tuning strategies. In particular, MaskSurf outperforms its closest competitor, Point-MAE, by 1.2\% on the real-world dataset of ScanObjectNN under the OBJ-BG setting, justifying the advantages of masked surfel prediction over masked point cloud reconstruction. Codes will be available at https://github.com/YBZh/MaskSurf.

Yabin Zhang, Bin Deng, Ruihuang Li et al. · stanford

It is well-known that the performance of well-trained deep neural networks may degrade significantly when they are applied to data with even slightly shifted distributions. Recent studies have shown that introducing certain perturbation on feature statistics (\eg, mean and standard deviation) during training can enhance the cross-domain generalization ability. Existing methods typically conduct such perturbation by utilizing the feature statistics within a mini-batch, limiting their representation capability. Inspired by the domain generalization objective, we introduce a novel Adversarial Style Augmentation (ASA) method, which explores broader style spaces by generating more effective statistics perturbation via adversarial training. Specifically, we first search for the most sensitive direction and intensity for statistics perturbation by maximizing the task loss. By updating the model against the adversarial statistics perturbation during training, we allow the model to explore the worst-case domain and hence improve its generalization performance. To facilitate the application of ASA, we design a simple yet effective module, namely AdvStyle, which instantiates the ASA method in a plug-and-play manner. We justify the efficacy of AdvStyle on tasks of cross-domain classification and instance retrieval. It achieves higher mean accuracy and lower performance fluctuation. Especially, our method significantly outperforms its competitors on the PACS dataset under the single source generalization setting, \eg, boosting the classification accuracy from 61.2\% to 67.1\% with a ResNet50 backbone. Our code will be available at \url{https://github.com/YBZh/AdvStyle}.

Yabin Zhang, Jiehong Lin, Ruihuang Li et al. · stanford

Masked autoencoder has demonstrated its effectiveness in self-supervised point cloud learning. Considering that masking is a kind of corruption, in this work we explore a more general denoising autoencoder for point cloud learning (Point-DAE) by investigating more types of corruptions beyond masking. Specifically, we degrade the point cloud with certain corruptions as input, and learn an encoder-decoder model to reconstruct the original point cloud from its corrupted version. Three corruption families (\ie, density/masking, noise, and affine transformation) and a total of fourteen corruption types are investigated with traditional non-Transformer encoders. Besides the popular masking corruption, we identify another effective corruption family, \ie, affine transformation. The affine transformation disturbs all points globally, which is complementary to the masking corruption where some local regions are dropped. We also validate the effectiveness of affine transformation corruption with the Transformer backbones, where we decompose the reconstruction of the complete point cloud into the reconstructions of detailed local patches and rough global shape, alleviating the position leakage problem in the reconstruction. Extensive experiments on tasks of object classification, few-shot learning, robustness testing, part segmentation, and 3D object detection validate the effectiveness of the proposed method. The codes are available at \url{https://github.com/YBZh/Point-DAE}.

Chenhang He, Ruihuang Li, Yabin Zhang et al. · stanford

Point cloud sequences are commonly used to accurately detect 3D objects in applications such as autonomous driving. Current top-performing multi-frame detectors mostly follow a Detect-and-Fuse framework, which extracts features from each frame of the sequence and fuses them to detect the objects in the current frame. However, this inevitably leads to redundant computation since adjacent frames are highly correlated. In this paper, we propose an efficient Motion-guided Sequential Fusion (MSF) method, which exploits the continuity of object motion to mine useful sequential contexts for object detection in the current frame. We first generate 3D proposals on the current frame and propagate them to preceding frames based on the estimated velocities. The points-of-interest are then pooled from the sequence and encoded as proposal features. A novel Bidirectional Feature Aggregation (BiFA) module is further proposed to facilitate the interactions of proposal features across frames. Besides, we optimize the point cloud pooling by a voxel-based sampling technique so that millions of points can be processed in several milliseconds. The proposed MSF method achieves not only better efficiency than other multi-frame detectors but also leading accuracy, with 83.12% and 78.30% mAP on the LEVEL1 and LEVEL2 test sets of Waymo Open Dataset, respectively. Codes can be found at \url{https://github.com/skyhehe123/MSF}.

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Minghan Li, Lei Zhang · microsoft-research

Many video instance segmentation (VIS) methods partition a video sequence into individual frames to detect and segment objects frame by frame. However, such a frame-in frame-out (FiFo) pipeline is ineffective to exploit the temporal information. Based on the fact that adjacent frames in a short clip are highly coherent in content, we propose to extend the one-stage FiFo framework to a clip-in clip-out (CiCo) one, which performs VIS clip by clip. Specifically, we stack FPN features of all frames in a short video clip to build a spatio-temporal feature cube, and replace the 2D conv layers in the prediction heads and the mask branch with 3D conv layers, forming clip-level prediction heads (CPH) and clip-level mask heads (CMH). Then the clip-level masks of an instance can be generated by feeding its box-level predictions from CPH and clip-level features from CMH into a small fully convolutional network. A clip-level segmentation loss is proposed to ensure that the generated instance masks are temporally coherent in the clip. The proposed CiCo strategy is free of inter-frame alignment, and can be easily embedded into existing FiFo based VIS approaches. To validate the generality and effectiveness of our CiCo strategy, we apply it to two representative FiFo methods, Yolact \cite{bolya2019yolact} and CondInst \cite{tian2020conditional}, resulting in two new one-stage VIS models, namely CiCo-Yolact and CiCo-CondInst, which achieve 37.1/37.3\%, 35.2/35.4\% and 17.2/18.0\% mask AP using the ResNet50 backbone, and 41.8/41.4\%, 38.0/38.9\% and 18.0/18.2\% mask AP using the Swin Transformer tiny backbone on YouTube-VIS 2019, 2021 and OVIS valid sets, respectively, recording new state-of-the-arts. Code and video demos of CiCo can be found at \url{https://github.com/MinghanLi/CiCo}.

Wangmeng Xiang, Chao Li, Yuxuan Zhou et al.

Skeleton-based action recognition has recently received considerable attention. Current approaches to skeleton-based action recognition are typically formulated as one-hot classification tasks and do not fully exploit the semantic relations between actions. For example, "make victory sign" and "thumb up" are two actions of hand gestures, whose major difference lies in the movement of hands. This information is agnostic from the categorical one-hot encoding of action classes but could be unveiled from the action description. Therefore, utilizing action description in training could potentially benefit representation learning. In this work, we propose a Generative Action-description Prompts (GAP) approach for skeleton-based action recognition. More specifically, we employ a pre-trained large-scale language model as the knowledge engine to automatically generate text descriptions for body parts movements of actions, and propose a multi-modal training scheme by utilizing the text encoder to generate feature vectors for different body parts and supervise the skeleton encoder for action representation learning. Experiments show that our proposed GAP method achieves noticeable improvements over various baseline models without extra computation cost at inference. GAP achieves new state-of-the-arts on popular skeleton-based action recognition benchmarks, including NTU RGB+D, NTU RGB+D 120 and NW-UCLA. The source code is available at https://github.com/MartinXM/GAP.

Feng Li, Ailing Zeng, Shilong Liu et al.

Recent DEtection TRansformer-based (DETR) models have obtained remarkable performance. Its success cannot be achieved without the re-introduction of multi-scale feature fusion in the encoder. However, the excessively increased tokens in multi-scale features, especially for about 75\% of low-level features, are quite computationally inefficient, which hinders real applications of DETR models. In this paper, we present Lite DETR, a simple yet efficient end-to-end object detection framework that can effectively reduce the GFLOPs of the detection head by 60\% while keeping 99\% of the original performance. Specifically, we design an efficient encoder block to update high-level features (corresponding to small-resolution feature maps) and low-level features (corresponding to large-resolution feature maps) in an interleaved way. In addition, to better fuse cross-scale features, we develop a key-aware deformable attention to predict more reliable attention weights. Comprehensive experiments validate the effectiveness and efficiency of the proposed Lite DETR, and the efficient encoder strategy can generalize well across existing DETR-based models. The code will be available in \url{https://github.com/IDEA-Research/Lite-DETR}.

Liqi Xue, Tianyi Xu, Yongbao Song et al. · microsoft-research

The display devices like HDR10 televisions are increasingly prevalent in our daily life for visualizing high dynamic range (HDR) images. But the majority of media images on the internet remain in 8-bit standard dynamic range (SDR) format. Therefore, converting SDR images to HDR ones by inverse tone mapping (ITM) is crucial to unlock the full potential of abundant media images. However, existing ITM methods are usually developed with complex network architectures requiring huge computational costs. In this paper, we propose a lightweight Improved Residual Network (IRNet) by enhancing the power of popular residual block for efficient ITM. Specifically, we propose a new Improved Residual Block (IRB) to extract and fuse multi-layer features for fine-grained HDR image reconstruction. Experiments on three benchmark datasets demonstrate that our IRNet achieves state-of-the-art performance on both the ITM and joint SR-ITM tasks. The code, models and data will be publicly available at https://github.com/ThisisVikki/ITM-baseline.

Haoxiang Zhang, Qixin Xu, Zhuofeng Li et al.

Long-horizon search agents accumulate large amounts of retrieved content across many tool calls, making context-budget efficiency increasingly important. A minimal intervention is to mask stale observations from the context as the trajectory progresses, but it remains unclear when this form of context management helps and why. We study observation masking through a systematic sweep over various agent backbones (4B to 284B parameters) and three retrievers on offline and live-web agentic search benchmarks. We find that the accuracy gain from masking follows an asymmetric inverted-U shape when plotted against the model's accuracy without context management: a plateau under weak retrievers, a peak when a strong retriever meets a mid-capacity model, and a sharp collapse when the model is saturated. This pattern reflects the interaction between retriever recall and the model's implicit filtering capacity, rather than either factor in isolation. Mechanistically, masking implements a token-for-turn trade-off: it removes observations the model has largely stopped attending to and pages the agent rarely re-opens. The added turns help when they convert failures into successes, but they fail when masking removes evidence the model would otherwise have used. We therefore reframe context management as a regime-dependent intervention and provide a holistic perspective for analyzing context use in agentic deep search. We release our scaffold and trajectories here (https://github.com/i-DeepSearch/observation-masking) to support future research.

Wangmeng Xiang, Chao Li, Biao Wang et al.

Transformer-based methods have recently achieved great advancement on 2D image-based vision tasks. For 3D video-based tasks such as action recognition, however, directly applying spatiotemporal transformers on video data will bring heavy computation and memory burdens due to the largely increased number of patches and the quadratic complexity of self-attention computation. How to efficiently and effectively model the 3D self-attention of video data has been a great challenge for transformers. In this paper, we propose a Temporal Patch Shift (TPS) method for efficient 3D self-attention modeling in transformers for video-based action recognition. TPS shifts part of patches with a specific mosaic pattern in the temporal dimension, thus converting a vanilla spatial self-attention operation to a spatiotemporal one with little additional cost. As a result, we can compute 3D self-attention using nearly the same computation and memory cost as 2D self-attention. TPS is a plug-and-play module and can be inserted into existing 2D transformer models to enhance spatiotemporal feature learning. The proposed method achieves competitive performance with state-of-the-arts on Something-something V1 & V2, Diving-48, and Kinetics400 while being much more efficient on computation and memory cost. The source code of TPS can be found at https://github.com/MartinXM/TPS.

Yukun Huang, Jianan Wang, Yukai Shi et al. · microsoft-research

Text-to-image diffusion models pre-trained on billions of image-text pairs have recently enabled 3D content creation by optimizing a randomly initialized differentiable 3D representation with score distillation. However, the optimization process suffers slow convergence and the resultant 3D models often exhibit two limitations: (a) quality concerns such as missing attributes and distorted shape and texture; (b) extremely low diversity comparing to text-guided image synthesis. In this paper, we show that the conflict between the 3D optimization process and uniform timestep sampling in score distillation is the main reason for these limitations. To resolve this conflict, we propose to prioritize timestep sampling with monotonically non-increasing functions, which aligns the 3D optimization process with the sampling process of diffusion model. Extensive experiments show that our simple redesign significantly improves 3D content creation with faster convergence, better quality and diversity.

Lei Zhang, Jie Zhang, Bowen Lei et al. · microsoft-research

Dataset Distillation (DD), a newly emerging field, aims at generating much smaller but efficient synthetic training datasets from large ones. Existing DD methods based on gradient matching achieve leading performance; however, they are extremely computationally intensive as they require continuously optimizing a dataset among thousands of randomly initialized models. In this paper, we assume that training the synthetic data with diverse models leads to better generalization performance. Thus we propose two model augmentation techniques, i.e. using early-stage models and parameter perturbation to learn an informative synthetic set with significantly reduced training cost. Extensive experiments demonstrate that our method achieves up to 20x speedup and comparable performance on par with state-of-the-art methods.

Ya-nan Guan, Shaonan Zhang, Hang Guo et al.

In this paper, we present a comprehensive overview of the NTIRE 2026 3rd Restore Any Image Model (RAIM) challenge, with a specific focus on Track 3: AI Flash Portrait. Despite significant advancements in deep learning for image restoration, existing models still encounter substantial challenges in real-world low-light portrait scenarios. Specifically, they struggle to achieve an optimal balance among noise suppression, detail preservation, and faithful illumination and color reproduction. To bridge this gap, this challenge aims to establish a novel benchmark for real-world low-light portrait restoration. We comprehensively evaluate the proposed algorithms utilizing a hybrid evaluation system that integrates objective quantitative metrics with rigorous subjective assessment protocols. For this competition, we provide a dataset containing 800 groups of real-captured low-light portrait data. Each group consists of a 1K-resolution low-light input image, a 1K ground truth (GT), and a 1K person mask. This challenge has garnered widespread attention from both academia and industry, attracting over 100 participating teams and receiving more than 3,000 valid submissions. This report details the motivation behind the challenge, the dataset construction process, the evaluation metrics, and the various phases of the competition. The released dataset and baseline code for this track are publicly available from the same \href{https://github.com/zsn1434/AI_Flash-BaseLine/tree/main}{GitHub repository}, and the official challenge webpage is hosted on \href{https://www.codabench.org/competitions/12885/}{CodaBench}.

Hao Zhang, Feng Li, Huaizhe Xu et al.

We present a mask-piloted Transformer which improves masked-attention in Mask2Former for image segmentation. The improvement is based on our observation that Mask2Former suffers from inconsistent mask predictions between consecutive decoder layers, which leads to inconsistent optimization goals and low utilization of decoder queries. To address this problem, we propose a mask-piloted training approach, which additionally feeds noised ground-truth masks in masked-attention and trains the model to reconstruct the original ones. Compared with the predicted masks used in mask-attention, the ground-truth masks serve as a pilot and effectively alleviate the negative impact of inaccurate mask predictions in Mask2Former. Based on this technique, our \M achieves a remarkable performance improvement on all three image segmentation tasks (instance, panoptic, and semantic), yielding $+2.3$AP and $+1.6$mIoU on the Cityscapes instance and semantic segmentation tasks with a ResNet-50 backbone. Our method also significantly speeds up the training, outperforming Mask2Former with half of the number of training epochs on ADE20K with both a ResNet-50 and a Swin-L backbones. Moreover, our method only introduces little computation during training and no extra computation during inference. Our code will be released at \url{https://github.com/IDEA-Research/MP-Former}.

Chenhang He, Ruihuang Li, Shuai Li et al.

Transformer has demonstrated promising performance in many 2D vision tasks. However, it is cumbersome to compute the self-attention on large-scale point cloud data because point cloud is a long sequence and unevenly distributed in 3D space. To solve this issue, existing methods usually compute self-attention locally by grouping the points into clusters of the same size, or perform convolutional self-attention on a discretized representation. However, the former results in stochastic point dropout, while the latter typically has narrow attention fields. In this paper, we propose a novel voxel-based architecture, namely Voxel Set Transformer (VoxSeT), to detect 3D objects from point clouds by means of set-to-set translation. VoxSeT is built upon a voxel-based set attention (VSA) module, which reduces the self-attention in each voxel by two cross-attentions and models features in a hidden space induced by a group of latent codes. With the VSA module, VoxSeT can manage voxelized point clusters with arbitrary size in a wide range, and process them in parallel with linear complexity. The proposed VoxSeT integrates the high performance of transformer with the efficiency of voxel-based model, which can be used as a good alternative to the convolutional and point-based backbones. VoxSeT reports competitive results on the KITTI and Waymo detection benchmarks. The source codes can be found at \url{https://github.com/skyhehe123/VoxSeT}.

Ruihuang Li, Chenhang He, Yabin Zhang et al. · stanford

Weakly supervised instance segmentation using only bounding box annotations has recently attracted much research attention. Most of the current efforts leverage low-level image features as extra supervision without explicitly exploiting the high-level semantic information of the objects, which will become ineffective when the foreground objects have similar appearances to the background or other objects nearby. We propose a new box-supervised instance segmentation approach by developing a Semantic-aware Instance Mask (SIM) generation paradigm. Instead of heavily relying on local pair-wise affinities among neighboring pixels, we construct a group of category-wise feature centroids as prototypes to identify foreground objects and assign them semantic-level pseudo labels. Considering that the semantic-aware prototypes cannot distinguish different instances of the same semantics, we propose a self-correction mechanism to rectify the falsely activated regions while enhancing the correct ones. Furthermore, to handle the occlusions between objects, we tailor the Copy-Paste operation for the weakly-supervised instance segmentation task to augment challenging training data. Extensive experimental results demonstrate the superiority of our proposed SIM approach over other state-of-the-art methods. The source code: https://github.com/lslrh/SIM.

Yabin Zhang, Wenjie Zhu, Chenhang He et al. · stanford

Out-of-distribution (OOD) detection is crucial for model reliability, as it identifies samples from unknown classes and reduces errors due to unexpected inputs. Vision-Language Models (VLMs) such as CLIP are emerging as powerful tools for OOD detection by integrating multi-modal information. However, the practical application of such systems is challenged by manual prompt engineering, which demands domain expertise and is sensitive to linguistic nuances. In this paper, we introduce Label-driven Automated Prompt Tuning (LAPT), a novel approach to OOD detection that reduces the need for manual prompt engineering. We develop distribution-aware prompts with in-distribution (ID) class names and negative labels mined automatically. Training samples linked to these class labels are collected autonomously via image synthesis and retrieval methods, allowing for prompt learning without manual effort. We utilize a simple cross-entropy loss for prompt optimization, with cross-modal and cross-distribution mixing strategies to reduce image noise and explore the intermediate space between distributions, respectively. The LAPT framework operates autonomously, requiring only ID class names as input and eliminating the need for manual intervention. With extensive experiments, LAPT consistently outperforms manually crafted prompts, setting a new standard for OOD detection. Moreover, LAPT not only enhances the distinction between ID and OOD samples, but also improves the ID classification accuracy and strengthens the generalization robustness to covariate shifts, resulting in outstanding performance in challenging full-spectrum OOD detection tasks. Codes are available at \url{https://github.com/YBZh/LAPT}.

Ailing Zeng, Muxi Chen, Lei Zhang et al.

Recently, there has been a surge of Transformer-based solutions for the long-term time series forecasting (LTSF) task. Despite the growing performance over the past few years, we question the validity of this line of research in this work. Specifically, Transformers is arguably the most successful solution to extract the semantic correlations among the elements in a long sequence. However, in time series modeling, we are to extract the temporal relations in an ordered set of continuous points. While employing positional encoding and using tokens to embed sub-series in Transformers facilitate preserving some ordering information, the nature of the \emph{permutation-invariant} self-attention mechanism inevitably results in temporal information loss. To validate our claim, we introduce a set of embarrassingly simple one-layer linear models named LTSF-Linear for comparison. Experimental results on nine real-life datasets show that LTSF-Linear surprisingly outperforms existing sophisticated Transformer-based LTSF models in all cases, and often by a large margin. Moreover, we conduct comprehensive empirical studies to explore the impacts of various design elements of LTSF models on their temporal relation extraction capability. We hope this surprising finding opens up new research directions for the LTSF task. We also advocate revisiting the validity of Transformer-based solutions for other time series analysis tasks (e.g., anomaly detection) in the future. Code is available at: \url{https://github.com/cure-lab/LTSF-Linear}.

Shuai Li, Chenhang He, Ruihuang Li et al.

Label assignment (LA), which aims to assign each training sample a positive (pos) and a negative (neg) loss weight, plays an important role in object detection. Existing LA methods mostly focus on the design of pos weighting function, while the neg weight is directly derived from the pos weight. Such a mechanism limits the learning capacity of detectors. In this paper, we explore a new weighting paradigm, termed dual weighting (DW), to specify pos and neg weights separately. We first identify the key influential factors of pos/neg weights by analyzing the evaluation metrics in object detection, and then design the pos and neg weighting functions based on them. Specifically, the pos weight of a sample is determined by the consistency degree between its classification and localization scores, while the neg weight is decomposed into two terms: the probability that it is a neg sample and its importance conditioned on being a neg sample. Such a weighting strategy offers greater flexibility to distinguish between important and less important samples, resulting in a more effective object detector. Equipped with the proposed DW method, a single FCOS-ResNet-50 detector can reach 41.5% mAP on COCO under 1x schedule, outperforming other existing LA methods. It consistently improves the baselines on COCO by a large margin under various backbones without bells and whistles. Code is available at https://github.com/strongwolf/DW.

Jie Liang, Hui Zeng, Lei Zhang

Single image super-resolution (SISR) with generative adversarial networks (GAN) has recently attracted increasing attention due to its potentials to generate rich details. However, the training of GAN is unstable, and it often introduces many perceptually unpleasant artifacts along with the generated details. In this paper, we demonstrate that it is possible to train a GAN-based SISR model which can stably generate perceptually realistic details while inhibiting visual artifacts. Based on the observation that the local statistics (e.g., residual variance) of artifact areas are often different from the areas of perceptually friendly details, we develop a framework to discriminate between GAN-generated artifacts and realistic details, and consequently generate an artifact map to regularize and stabilize the model training process. Our proposed locally discriminative learning (LDL) method is simple yet effective, which can be easily plugged in off-the-shelf SISR methods and boost their performance. Experiments demonstrate that LDL outperforms the state-of-the-art GAN based SISR methods, achieving not only higher reconstruction accuracy but also superior perceptual quality on both synthetic and real-world datasets. Codes and models are available at https://github.com/csjliang/LDL.

Xiaoming Li, Chaofeng Chen, Xianhui Lin et al.

How to design proper training pairs is critical for super-resolving real-world low-quality (LQ) images, which suffers from the difficulties in either acquiring paired ground-truth high-quality (HQ) images or synthesizing photo-realistic degraded LQ observations. Recent works mainly focus on modeling the degradation with handcrafted or estimated degradation parameters, which are however incapable to model complicated real-world degradation types, resulting in limited quality improvement. Notably, LQ face images, which may have the same degradation process as natural images, can be robustly restored with photo-realistic textures by exploiting their strong structural priors. This motivates us to use the real-world LQ face images and their restored HQ counterparts to model the complex real-world degradation (namely ReDegNet), and then transfer it to HQ natural images to synthesize their realistic LQ counterparts. By taking these paired HQ-LQ face images as inputs to explicitly predict the degradation-aware and content-independent representations, we could control the degraded image generation, and subsequently transfer these degradation representations from face to natural images to synthesize the degraded LQ natural images. Experiments show that our ReDegNet can well learn the real degradation process from face images. The restoration network trained with our synthetic pairs performs favorably against SOTAs. More importantly, our method provides a new way to handle the real-world complex scenarios by learning their degradation representations from the facial portions, which can be used to significantly improve the quality of non-facial areas. The source code is available at https://github.com/csxmli2016/ReDegNet.

Shilong Liu, Hao Cheng, Haotian Liu et al. · microsoft-research

LLaVA-Plus is a general-purpose multimodal assistant that expands the capabilities of large multimodal models. It maintains a skill repository of pre-trained vision and vision-language models and can activate relevant tools based on users' inputs to fulfill real-world tasks. LLaVA-Plus is trained on multimodal instruction-following data to acquire the ability to use tools, covering visual understanding, generation, external knowledge retrieval, and compositions. Empirical results show that LLaVA-Plus outperforms LLaVA in existing capabilities and exhibits new ones. It is distinct in that the image query is directly grounded and actively engaged throughout the entire human-AI interaction sessions, significantly improving tool use performance and enabling new scenarios.

Ruihuang Li, Shuai Li, Chenhang He et al. · stanford

Domain adaptive semantic segmentation aims to learn a model with the supervision of source domain data, and produce satisfactory dense predictions on unlabeled target domain. One popular solution to this challenging task is self-training, which selects high-scoring predictions on target samples as pseudo labels for training. However, the produced pseudo labels often contain much noise because the model is biased to source domain as well as majority categories. To address the above issues, we propose to directly explore the intrinsic pixel distributions of target domain data, instead of heavily relying on the source domain. Specifically, we simultaneously cluster pixels and rectify pseudo labels with the obtained cluster assignments. This process is done in an online fashion so that pseudo labels could co-evolve with the segmentation model without extra training rounds. To overcome the class imbalance problem on long-tailed categories, we employ a distribution alignment technique to enforce the marginal class distribution of cluster assignments to be close to that of pseudo labels. The proposed method, namely Class-balanced Pixel-level Self-Labeling (CPSL), improves the segmentation performance on target domain over state-of-the-arts by a large margin, especially on long-tailed categories.

Rang Meng, Xianfeng Li, Weijie Chen et al.

Convolutional neural networks (CNNs) have demonstrated gratifying results at learning discriminative features. However, when applied to unseen domains, state-of-the-art models are usually prone to errors due to domain shift. After investigating this issue from the perspective of shortcut learning, we find the devils lie in the fact that models trained on different domains merely bias to different domain-specific features yet overlook diverse task-related features. Under this guidance, a novel Attention Diversification framework is proposed, in which Intra-Model and Inter-Model Attention Diversification Regularization are collaborated to reassign appropriate attention to diverse task-related features. Briefly, Intra-Model Attention Diversification Regularization is equipped on the high-level feature maps to achieve in-channel discrimination and cross-channel diversification via forcing different channels to pay their most salient attention to different spatial locations. Besides, Inter-Model Attention Diversification Regularization is proposed to further provide task-related attention diversification and domain-related attention suppression, which is a paradigm of "simulate, divide and assemble": simulate domain shift via exploiting multiple domain-specific models, divide attention maps into task-related and domain-related groups, and assemble them within each group respectively to execute regularization. Extensive experiments and analyses are conducted on various benchmarks to demonstrate that our method achieves state-of-the-art performance over other competing methods. Code is available at https://github.com/hikvision-research/DomainGeneralization.

Tao Chen, Yazhou Yao, Lei Zhang et al.

Weakly supervised semantic segmentation with only image-level labels aims to reduce annotation costs for the segmentation task. Existing approaches generally leverage class activation maps (CAMs) to locate the object regions for pseudo label generation. However, CAMs can only discover the most discriminative parts of objects, thus leading to inferior pixel-level pseudo labels. To address this issue, we propose a saliency guided Inter- and Intra-Class Relation Constrained (I$^2$CRC) framework to assist the expansion of the activated object regions in CAMs. Specifically, we propose a saliency guided class-agnostic distance module to pull the intra-category features closer by aligning features to their class prototypes. Further, we propose a class-specific distance module to push the inter-class features apart and encourage the object region to have a higher activation than the background. Besides strengthening the capability of the classification network to activate more integral object regions in CAMs, we also introduce an object guided label refinement module to take a full use of both the segmentation prediction and the initial labels for obtaining superior pseudo-labels. Extensive experiments on PASCAL VOC 2012 and COCO datasets demonstrate well the effectiveness of I$^2$CRC over other state-of-the-art counterparts. The source codes, models, and data have been made available at \url{https://github.com/NUST-Machine-Intelligence-Laboratory/I2CRC}.

Jie Liang, Hui Zeng, Lei Zhang

Efficient and effective real-world image super-resolution (Real-ISR) is a challenging task due to the unknown complex degradation of real-world images and the limited computation resources in practical applications. Recent research on Real-ISR has achieved significant progress by modeling the image degradation space; however, these methods largely rely on heavy backbone networks and they are inflexible to handle images of different degradation levels. In this paper, we propose an efficient and effective degradation-adaptive super-resolution (DASR) network, whose parameters are adaptively specified by estimating the degradation of each input image. Specifically, a tiny regression network is employed to predict the degradation parameters of the input image, while several convolutional experts with the same topology are jointly optimized to specify the network parameters via a non-linear mixture of experts. The joint optimization of multiple experts and the degradation-adaptive pipeline significantly extend the model capacity to handle degradations of various levels, while the inference remains efficient since only one adaptively specified network is used for super-resolving the input image. Our extensive experiments demonstrate that the proposed DASR is not only much more effective than existing methods on handling real-world images with different degradation levels but also efficient for easy deployment. Codes, models and datasets are available at https://github.com/csjliang/DASR.

Lishen Qu, Yao Liu, Jie Liang et al.

This paper presents NTIRE 2026, the 3rd Restore Any Image Model (RAIM) challenge on multi-exposure image fusion in dynamic scenes. We introduce a benchmark that targets a practical yet difficult HDR imaging setting, where exposure bracketing must be fused under scene motion, illumination variation, and handheld camera jitter. The challenge data contains 100 training sequences with 7 exposure levels and 100 test sequences with 5 exposure levels, reflecting real-world scenarios that frequently cause misalignment and ghosting artefacts. We evaluate submissions with a leaderboard score derived from PSNR, SSIM, and LPIPS, while also considering perceptual quality, efficiency, and reproducibility during the final review. This track attracted 114 participating teams and received 987 submissions. The winning methods significantly improved the ability to remove artifacts from multi-exposure fusion and recover fine details. The dataset and the code of each team can be found at the repository: https://github.com/qulishen/RAIM-HDR.

Tianhe Ren, Shilong Liu, Feng Li et al.

The DEtection TRansformer (DETR) algorithm has received considerable attention in the research community and is gradually emerging as a mainstream approach for object detection and other perception tasks. However, the current field lacks a unified and comprehensive benchmark specifically tailored for DETR-based models. To address this issue, we develop a unified, highly modular, and lightweight codebase called detrex, which supports a majority of the mainstream DETR-based instance recognition algorithms, covering various fundamental tasks, including object detection, segmentation, and pose estimation. We conduct extensive experiments under detrex and perform a comprehensive benchmark for DETR-based models. Moreover, we enhance the performance of detection transformers through the refinement of training hyper-parameters, providing strong baselines for supported algorithms.We hope that detrex could offer research communities a standardized and unified platform to evaluate and compare different DETR-based models while fostering a deeper understanding and driving advancements in DETR-based instance recognition. Our code is available at https://github.com/IDEA-Research/detrex. The project is currently being actively developed. We encourage the community to use detrex codebase for further development and contributions.

Lei Zhang, Yuge Zhang, Kan Ren et al.

The field of machine learning (ML) has gained widespread adoption, leading to significant demand for adapting ML to specific scenarios, which is yet expensive and non-trivial. The predominant approaches towards the automation of solving ML tasks (e.g., AutoML) are often time-consuming and hard to understand for human developers. In contrast, though human engineers have the incredible ability to understand tasks and reason about solutions, their experience and knowledge are often sparse and difficult to utilize by quantitative approaches. In this paper, we aim to bridge the gap between machine intelligence and human knowledge by introducing a novel framework, which leverages the state-of-the-art large language models to develop ML solutions for novel tasks. We showcase the possibility of extending the capability of LLMs to comprehend structured inputs and perform thorough reasoning for solving novel ML tasks. And we find that, after some dedicated design, the LLM can (i) observe from the existing experiences of ML tasks and (ii) reason effectively to deliver promising results for new tasks. The solution generated can be used directly to achieve high levels of competitiveness. Examples and code available at https://github.com/microsoft/CoML.

Shilong Liu, Tianhe Ren, Jiayu Chen et al.

This paper is concerned with the matching stability problem across different decoder layers in DEtection TRansformers (DETR). We point out that the unstable matching in DETR is caused by a multi-optimization path problem, which is highlighted by the one-to-one matching design in DETR. To address this problem, we show that the most important design is to use and only use positional metrics (like IOU) to supervise classification scores of positive examples. Under the principle, we propose two simple yet effective modifications by integrating positional metrics to DETR's classification loss and matching cost, named position-supervised loss and position-modulated cost. We verify our methods on several DETR variants. Our methods show consistent improvements over baselines. By integrating our methods with DINO, we achieve 50.4 and 51.5 AP on the COCO detection benchmark using ResNet-50 backbones under 12 epochs and 24 epochs training settings, achieving a new record under the same setting. We achieve 63.8 AP on COCO detection test-dev with a Swin-Large backbone. Our code will be made available at https://github.com/IDEA-Research/Stable-DINO.

Tao Yang, Rongyuan Wu, Peiran Ren et al.

Diffusion models have demonstrated impressive performance in various image generation, editing, enhancement and translation tasks. In particular, the pre-trained text-to-image stable diffusion models provide a potential solution to the challenging realistic image super-resolution (Real-ISR) and image stylization problems with their strong generative priors. However, the existing methods along this line often fail to keep faithful pixel-wise image structures. If extra skip connections between the encoder and the decoder of a VAE are used to reproduce details, additional training in image space will be required, limiting the application to tasks in latent space such as image stylization. In this work, we propose a pixel-aware stable diffusion (PASD) network to achieve robust Real-ISR and personalized image stylization. Specifically, a pixel-aware cross attention module is introduced to enable diffusion models perceiving image local structures in pixel-wise level, while a degradation removal module is used to extract degradation insensitive features to guide the diffusion process together with image high level information. An adjustable noise schedule is introduced to further improve the image restoration results. By simply replacing the base diffusion model with a stylized one, PASD can generate diverse stylized images without collecting pairwise training data, and by shifting the base model with an aesthetic one, PASD can bring old photos back to life. Extensive experiments in a variety of image enhancement and stylization tasks demonstrate the effectiveness of our proposed PASD approach. Our source codes are available at \url{https://github.com/yangxy/PASD/}.

Xindong Zhang, Hui Zeng, Shi Guo et al.

Recently, transformer-based methods have demonstrated impressive results in various vision tasks, including image super-resolution (SR), by exploiting the self-attention (SA) for feature extraction. However, the computation of SA in most existing transformer based models is very expensive, while some employed operations may be redundant for the SR task. This limits the range of SA computation and consequently the SR performance. In this work, we propose an efficient long-range attention network (ELAN) for image SR. Specifically, we first employ shift convolution (shift-conv) to effectively extract the image local structural information while maintaining the same level of complexity as 1x1 convolution, then propose a group-wise multi-scale self-attention (GMSA) module, which calculates SA on non-overlapped groups of features using different window sizes to exploit the long-range image dependency. A highly efficient long-range attention block (ELAB) is then built by simply cascading two shift-conv with a GMSA module, which is further accelerated by using a shared attention mechanism. Without bells and whistles, our ELAN follows a fairly simple design by sequentially cascading the ELABs. Extensive experiments demonstrate that ELAN obtains even better results against the transformer-based SR models but with significantly less complexity. The source code can be found at https://github.com/xindongzhang/ELAN.

Xianbiao Qi, Jianan Wang, Yihao Chen et al.

We present a Lipschitz continuous Transformer, called LipsFormer, to pursue training stability both theoretically and empirically for Transformer-based models. In contrast to previous practical tricks that address training instability by learning rate warmup, layer normalization, attention formulation, and weight initialization, we show that Lipschitz continuity is a more essential property to ensure training stability. In LipsFormer, we replace unstable Transformer component modules with Lipschitz continuous counterparts: CenterNorm instead of LayerNorm, spectral initialization instead of Xavier initialization, scaled cosine similarity attention instead of dot-product attention, and weighted residual shortcut. We prove that these introduced modules are Lipschitz continuous and derive an upper bound on the Lipschitz constant of LipsFormer. Our experiments show that LipsFormer allows stable training of deep Transformer architectures without the need of careful learning rate tuning such as warmup, yielding a faster convergence and better generalization. As a result, on the ImageNet 1K dataset, LipsFormer-Swin-Tiny based on Swin Transformer training for 300 epochs can obtain 82.7\% without any learning rate warmup. Moreover, LipsFormer-CSwin-Tiny, based on CSwin, training for 300 epochs achieves a top-1 accuracy of 83.5\% with 4.7G FLOPs and 24M parameters. The code will be released at \url{https://github.com/IDEA-Research/LipsFormer}.

Wentong Li, Wenyu Liu, Jianke Zhu et al.

In contrast to the fully supervised methods using pixel-wise mask labels, box-supervised instance segmentation takes advantage of the simple box annotations, which has recently attracted a lot of research attentions. In this paper, we propose a novel single-shot box-supervised instance segmentation approach, which integrates the classical level set model with deep neural network delicately. Specifically, our proposed method iteratively learns a series of level sets through a continuous Chan-Vese energy-based function in an end-to-end fashion. A simple mask supervised SOLOv2 model is adapted to predict the instance-aware mask map as the level set for each instance. Both the input image and its deep features are employed as the input data to evolve the level set curves, where a box projection function is employed to obtain the initial boundary. By minimizing the fully differentiable energy function, the level set for each instance is iteratively optimized within its corresponding bounding box annotation. The experimental results on four challenging benchmarks demonstrate the leading performance of our proposed approach to robust instance segmentation in various scenarios. The code is available at: https://github.com/LiWentomng/boxlevelset.

Pandeng Li, Chen-Wei Xie, Hongtao Xie et al.

Video moment retrieval pursues an efficient and generalized solution to identify the specific temporal segments within an untrimmed video that correspond to a given language description. To achieve this goal, we provide a generative diffusion-based framework called MomentDiff, which simulates a typical human retrieval process from random browsing to gradual localization. Specifically, we first diffuse the real span to random noise, and learn to denoise the random noise to the original span with the guidance of similarity between text and video. This allows the model to learn a mapping from arbitrary random locations to real moments, enabling the ability to locate segments from random initialization. Once trained, MomentDiff could sample random temporal segments as initial guesses and iteratively refine them to generate an accurate temporal boundary. Different from discriminative works (e.g., based on learnable proposals or queries), MomentDiff with random initialized spans could resist the temporal location biases from datasets. To evaluate the influence of the temporal location biases, we propose two anti-bias datasets with location distribution shifts, named Charades-STA-Len and Charades-STA-Mom. The experimental results demonstrate that our efficient framework consistently outperforms state-of-the-art methods on three public benchmarks, and exhibits better generalization and robustness on the proposed anti-bias datasets. The code, model, and anti-bias evaluation datasets are available at https://github.com/IMCCretrieval/MomentDiff.

Shilong Liu, Yaoyuan Liang, Feng Li et al.

In this paper, we study the problem of visual grounding by considering both phrase extraction and grounding (PEG). In contrast to the previous phrase-known-at-test setting, PEG requires a model to extract phrases from text and locate objects from images simultaneously, which is a more practical setting in real applications. As phrase extraction can be regarded as a $1$D text segmentation problem, we formulate PEG as a dual detection problem and propose a novel DQ-DETR model, which introduces dual queries to probe different features from image and text for object prediction and phrase mask prediction. Each pair of dual queries is designed to have shared positional parts but different content parts. Such a design effectively alleviates the difficulty of modality alignment between image and text (in contrast to a single query design) and empowers Transformer decoder to leverage phrase mask-guided attention to improve performance. To evaluate the performance of PEG, we also propose a new metric CMAP (cross-modal average precision), analogous to the AP metric in object detection. The new metric overcomes the ambiguity of Recall@1 in many-box-to-one-phrase cases in phrase grounding. As a result, our PEG pre-trained DQ-DETR establishes new state-of-the-art results on all visual grounding benchmarks with a ResNet-101 backbone. For example, it achieves $91.04\%$ and $83.51\%$ in terms of recall rate on RefCOCO testA and testB with a ResNet-101 backbone. Code will be availabl at \url{https://github.com/IDEA-Research/DQ-DETR}.

Lei Zhang, Guanyu Gao, Huaizheng Zhang

Data drift is a thorny challenge when deploying person re-identification (ReID) models into real-world devices, where the data distribution is significantly different from that of the training environment and keeps changing. To tackle this issue, we propose a federated spatial-temporal incremental learning approach, named FedSTIL, which leverages both lifelong learning and federated learning to continuously optimize models deployed on many distributed edge clients. Unlike previous efforts, FedSTIL aims to mine spatial-temporal correlations among the knowledge learnt from different edge clients. Specifically, the edge clients first periodically extract general representations of drifted data to optimize their local models. Then, the learnt knowledge from edge clients will be aggregated by centralized parameter server, where the knowledge will be selectively and attentively distilled from spatial- and temporal-dimension with carefully designed mechanisms. Finally, the distilled informative spatial-temporal knowledge will be sent back to correlated edge clients to further improve the recognition accuracy of each edge client with a lifelong learning method. Extensive experiments on a mixture of five real-world datasets demonstrate that our method outperforms others by nearly 4% in Rank-1 accuracy, while reducing communication cost by 62%. All implementation codes are publicly available on https://github.com/MSNLAB/Federated-Lifelong-Person-ReID

Lei Zhang, Xiaoke Wang, Michael Rawson et al.

Purpose To develop and evaluate a deep learning-based method (MC-Net) to suppress motion artifacts in brain magnetic resonance imaging (MRI). Methods MC-Net was derived from a UNet combined with a two-stage multi-loss function. T1-weighted axial brain images contaminated with synthetic motions were used to train the network. Evaluation used simulated T1 and T2-weighted axial, coronal, and sagittal images unseen during training, as well as T1-weighted images with motion artifacts from real scans. Performance indices included the peak signal to noise ratio (PSNR), structural similarity index measure (SSIM), and visual reading scores. Two clinical readers scored the images. Results The MC-Net outperformed other methods implemented in terms of PSNR and SSIM on the T1 axial test set. The MC-Net significantly improved the quality of all T1-weighted images (for all directions and for simulated as well as real motion artifacts), both on quantitative measures and visual scores. However, the MC-Net performed poorly on images of untrained contrast (T2-weighted). Conclusion The proposed two-stage multi-loss MC-Net can effectively suppress motion artifacts in brain MRI without compromising image context. Given the efficiency of the MC-Net (single image processing time ~40ms), it can potentially be used in real clinical settings. To facilitate further research, the code and trained model are available at https://github.com/MRIMoCo/DL_Motion_Correction.