Wenjie Pei

Qi Fan, Wenjie Pei, Yu-Wing Tai et al.

Existing few-shot segmentation methods have achieved great progress based on the support-query matching framework. But they still heavily suffer from the limited coverage of intra-class variations from the few-shot supports provided. Motivated by the simple Gestalt principle that pixels belonging to the same object are more similar than those to different objects of same class, we propose a novel self-support matching strategy to alleviate this problem, which uses query prototypes to match query features, where the query prototypes are collected from high-confidence query predictions. This strategy can effectively capture the consistent underlying characteristics of the query objects, and thus fittingly match query features. We also propose an adaptive self-support background prototype generation module and self-support loss to further facilitate the self-support matching procedure. Our self-support network substantially improves the prototype quality, benefits more improvement from stronger backbones and more supports, and achieves SOTA on multiple datasets. Codes are at \url{https://github.com/fanq15/SSP}.

Jing Xu, Xu Luo, Xinglin Pan et al.

Few-shot learning (FSL) targets at generalization of vision models towards unseen tasks without sufficient annotations. Despite the emergence of a number of few-shot learning methods, the sample selection bias problem, i.e., the sensitivity to the limited amount of support data, has not been well understood. In this paper, we find that this problem usually occurs when the positions of support samples are in the vicinity of task centroid -- the mean of all class centroids in the task. This motivates us to propose an extremely simple feature transformation to alleviate this problem, dubbed Task Centroid Projection Removing (TCPR). TCPR is applied directly to all image features in a given task, aiming at removing the dimension of features along the direction of the task centroid. While the exact task centroid cannot be accurately obtained from limited data, we estimate it using base features that are each similar to one of the support features. Our method effectively prevents features from being too close to the task centroid. Extensive experiments over ten datasets from different domains show that TCPR can reliably improve classification accuracy across various feature extractors, training algorithms and datasets. The code has been made available at https://github.com/KikimorMay/FSL-TCBR.

Zhenhua Ning, Zhuotao Tian, Guangming Lu et al.

Although extensive research has been conducted on 3D point cloud segmentation, effectively adapting generic models to novel categories remains a formidable challenge. This paper proposes a novel approach to improve point cloud few-shot segmentation (PC-FSS) models. Unlike existing PC-FSS methods that directly utilize categorical information from support prototypes to recognize novel classes in query samples, our method identifies two critical aspects that substantially enhance model performance by reducing contextual gaps between support prototypes and query features. Specifically, we (1) adapt support background prototypes to match query context while removing extraneous cues that may obscure foreground and background in query samples, and (2) holistically rectify support prototypes under the guidance of query features to emulate the latter having no semantic gap to the query targets. Our proposed designs are agnostic to the feature extractor, rendering them readily applicable to any prototype-based methods. The experimental results on S3DIS and ScanNet demonstrate notable practical benefits, as our approach achieves significant improvements while still maintaining high efficiency. The code for our approach is available at https://github.com/AaronNZH/Boosting-Few-shot-3D-Point-Cloud-Segmentation-via-Query-Guided-Enhancement

Xin Feng, Yifeng Xu, Guangming Lu et al.

Effective image restoration with large-size corruptions, such as blind image inpainting, entails precise detection of corruption region masks which remains extremely challenging due to diverse shapes and patterns of corruptions. In this work, we present a novel method for automatic corruption detection, which allows for blind corruption restoration without known corruption masks. Specifically, we develop a hierarchical contrastive learning framework to detect corrupted regions by capturing the intrinsic semantic distinctions between corrupted and uncorrupted regions. In particular, our model detects the corrupted mask in a coarse-to-fine manner by first predicting a coarse mask by contrastive learning in low-resolution feature space and then refines the uncertain area of the mask by high-resolution contrastive learning. A specialized hierarchical interaction mechanism is designed to facilitate the knowledge propagation of contrastive learning in different scales, boosting the modeling performance substantially. The detected multi-scale corruption masks are then leveraged to guide the corruption restoration. Detecting corrupted regions by learning the contrastive distinctions rather than the semantic patterns of corruptions, our model has well generalization ability across different corruption patterns. Extensive experiments demonstrate following merits of our model: 1) the superior performance over other methods on both corruption detection and various image restoration tasks including blind inpainting and watermark removal, and 2) strong generalization across different corruption patterns such as graffiti, random noise or other image content. Codes and trained weights are available at https://github.com/xyfJASON/HCL .

Zhengyao Fang, Pengyuan Lyu, Chengquan Zhang et al.

Vision-language models (VLMs) face significant computational inefficiencies caused by excessive generation of visual tokens. While prior work shows that a large fraction of visual tokens are redundant, existing compression methods struggle to balance importance preservation and information diversity. To address this, we propose PruneSID, a training-free Synergistic Importance-Diversity approach featuring a two-stage pipeline: (1) Principal Semantic Components Analysis (PSCA) for clustering tokens into semantically coherent groups, ensuring comprehensive concept coverage, and (2) Intra-group Non-Maximum Suppression (NMS) for pruning redundant tokens while preserving key representative tokens within each group. Additionally, PruneSID incorporates an information-aware dynamic compression ratio mechanism that optimizes token compression rates based on image complexity, enabling more effective average information preservation across diverse scenes. Extensive experiments demonstrate state-of-the-art performance, achieving 96.3% accuracy on LLaVA-1.5 with only 11.1% token retention, and 92.8% accuracy at extreme compression rates (5.6%) on LLaVA-NeXT, outperforming prior methods by 2.5% with 7.8 $\times$ faster prefilling speed compared to the original model. Our framework generalizes across diverse VLMs and both image and video modalities, showcasing strong cross-modal versatility. Code is available at https://github.com/ZhengyaoFang/PruneSID.

Zhengyao Fang, Zexi Jia, Yijia Zhong et al.

Recent advances in text-to-image (T2I) generation have greatly improved visual quality, yet producing images that appear visually authentic to real-world photography remains challenging. This is partly due to biases in existing evaluation paradigms: human ratings and preference-trained metrics often favor visually vivid images with exaggerated saturation and contrast, which make generations often too vivid to be real even when prompted for realistic-style images. To address this issue, we present Color Fidelity Dataset (CFD) and Color Fidelity Metric (CFM) for objective evaluation of color fidelity in realistic-style generations. CFD contains over 1.3M real and synthetic images with ordered levels of color realism, while CFM employs a multimodal encoder to learn perceptual color fidelity. In addition, we propose a training-free Color Fidelity Refinement (CFR) that adaptively modulates spatial-temporal guidance scale in generation, thereby enhancing color authenticity. Together, CFD supports CFM for assessment, whose learned attention further guides CFR to refine T2I fidelity, forming a progressive framework for assessing and improving color fidelity in realistic-style T2I generation. The dataset and code are available at https://github.com/ZhengyaoFang/CFM.

Shuang Wu, Songlin Tang, Guangming Lu et al.

Typical inverse rendering methods focus on learning implicit neural scene representations by modeling the geometry, materials and illumination separately, which entails significant computations for optimization. In this work we design a Unified Voxelization framework for explicit learning of scene representations, dubbed UniVoxel, which allows for efficient modeling of the geometry, materials and illumination jointly, thereby accelerating the inverse rendering significantly. To be specific, we propose to encode a scene into a latent volumetric representation, based on which the geometry, materials and illumination can be readily learned via lightweight neural networks in a unified manner. Particularly, an essential design of UniVoxel is that we leverage local Spherical Gaussians to represent the incident light radiance, which enables the seamless integration of modeling illumination into the unified voxelization framework. Such novel design enables our UniVoxel to model the joint effects of direct lighting, indirect lighting and light visibility efficiently without expensive multi-bounce ray tracing. Extensive experiments on multiple benchmarks covering diverse scenes demonstrate that UniVoxel boosts the optimization efficiency significantly compared to other methods, reducing the per-scene training time from hours to 18 minutes, while achieving favorable reconstruction quality. Code is available at https://github.com/freemantom/UniVoxel.

Songlin Tang, Wenjie Pei, Xin Tao et al.

Existing methods for interactive segmentation in radiance fields entail scene-specific optimization and thus cannot generalize across different scenes, which greatly limits their applicability. In this work we make the first attempt at Scene-Generalizable Interactive Segmentation in Radiance Fields (SGISRF) and propose a novel SGISRF method, which can perform 3D object segmentation for novel (unseen) scenes represented by radiance fields, guided by only a few interactive user clicks in a given set of multi-view 2D images. In particular, the proposed SGISRF focuses on addressing three crucial challenges with three specially designed techniques. First, we devise the Cross-Dimension Guidance Propagation to encode the scarce 2D user clicks into informative 3D guidance representations. Second, the Uncertainty-Eliminated 3D Segmentation module is designed to achieve efficient yet effective 3D segmentation. Third, Concealment-Revealed Supervised Learning scheme is proposed to reveal and correct the concealed 3D segmentation errors resulted from the supervision in 2D space with only 2D mask annotations. Extensive experiments on two real-world challenging benchmarks covering diverse scenes demonstrate 1) effectiveness and scene-generalizability of the proposed method, 2) favorable performance compared to classical method requiring scene-specific optimization.

Huimin Zeng, Weinong Wang, Xin Tao et al.

Recent interactive segmentation methods iteratively take source image, user guidance and previously predicted mask as the input without considering the invariant nature of the source image. As a result, extracting features from the source image is repeated in each interaction, resulting in substantial computational redundancy. In this work, we propose the Feature Decoupling-Recycling Network (FDRN), which decouples the modeling components based on their intrinsic discrepancies and then recycles components for each user interaction. Thus, the efficiency of the whole interactive process can be significantly improved. To be specific, we apply the Decoupling-Recycling strategy from three perspectives to address three types of discrepancies, respectively. First, our model decouples the learning of source image semantics from the encoding of user guidance to process two types of input domains separately. Second, FDRN decouples high-level and low-level features from stratified semantic representations to enhance feature learning. Third, during the encoding of user guidance, current user guidance is decoupled from historical guidance to highlight the effect of current user guidance. We conduct extensive experiments on 6 datasets from different domains and modalities, which demonstrate the following merits of our model: 1) superior efficiency than other methods, particularly advantageous in challenging scenarios requiring long-term interactions (up to 4.25x faster), while achieving favorable segmentation performance; 2) strong applicability to various methods serving as a universal enhancement technique; 3) well cross-task generalizability, e.g., to medical image segmentation, and robustness against misleading user guidance.

Shuang Wu, Wenjie Pei, Dianwen Mei et al.

Most of existing methods for few-shot object detection follow the fine-tuning paradigm, which potentially assumes that the class-agnostic generalizable knowledge can be learned and transferred implicitly from base classes with abundant samples to novel classes with limited samples via such a two-stage training strategy. However, it is not necessarily true since the object detector can hardly distinguish between class-agnostic knowledge and class-specific knowledge automatically without explicit modeling. In this work we propose to learn three types of class-agnostic commonalities between base and novel classes explicitly: recognition-related semantic commonalities, localization-related semantic commonalities and distribution commonalities. We design a unified distillation framework based on a memory bank, which is able to perform distillation of all three types of commonalities jointly and efficiently. Extensive experiments demonstrate that our method can be readily integrated into most of existing fine-tuning based methods and consistently improve the performance by a large margin.

Wenjie Pei, Xin Feng, Canmiao Fu et al.

The key challenge of sequence representation learning is to capture the long-range temporal dependencies. Typical methods for supervised sequence representation learning are built upon recurrent neural networks to capture temporal dependencies. One potential limitation of these methods is that they only model one-order information interactions explicitly between adjacent time steps in a sequence, hence the high-order interactions between nonadjacent time steps are not fully exploited. It greatly limits the capability of modeling the long-range temporal dependencies since the temporal features learned by one-order interactions cannot be maintained for a long term due to temporal information dilution and gradient vanishing. To tackle this limitation, we propose the Non-local Recurrent Neural Memory (NRNM) for supervised sequence representation learning, which performs non-local operations \MR{by means of self-attention mechanism} to learn full-order interactions within a sliding temporal memory block and models global interactions between memory blocks in a gated recurrent manner. Consequently, our model is able to capture long-range dependencies. Besides, the latent high-level features contained in high-order interactions can be distilled by our model. We validate the effectiveness and generalization of our NRNM on three types of sequence applications across different modalities, including sequence classification, step-wise sequential prediction and sequence similarity learning. Our model compares favorably against other state-of-the-art methods specifically designed for each of these sequence applications.

Xin Feng, Haobo Ji, Wenjie Pei et al.

While the research on image background restoration from regular size of degraded images has achieved remarkable progress, restoring ultra high-resolution (e.g., 4K) images remains an extremely challenging task due to the explosion of computational complexity and memory usage, as well as the deficiency of annotated data. In this paper we present a novel model for ultra high-resolution image restoration, referred to as the Global-Local Stepwise Generative Network (GLSGN), which employs a stepwise restoring strategy involving four restoring pathways: three local pathways and one global pathway. The local pathways focus on conducting image restoration in a fine-grained manner over local but high-resolution image patches, while the global pathway performs image restoration coarsely on the scale-down but intact image to provide cues for the local pathways in a global view including semantics and noise patterns. To smooth the mutual collaboration between these four pathways, our GLSGN is designed to ensure the inter-pathway consistency in four aspects in terms of low-level content, perceptual attention, restoring intensity and high-level semantics, respectively. As another major contribution of this work, we also introduce the first ultra high-resolution dataset to date for both reflection removal and rain streak removal, comprising 4,670 real-world and synthetic images. Extensive experiments across three typical tasks for image background restoration, including image reflection removal, image rain streak removal and image dehazing, show that our GLSGN consistently outperforms state-of-the-art methods.

Jiadong Liang, Wenjie Pei, Feng Lu

The crux of text-to-image synthesis stems from the difficulty of preserving the cross-modality semantic consistency between the input text and the synthesized image. Typical methods, which seek to model the text-to-image mapping directly, could only capture keywords in the text that indicates common objects or actions but fail to learn their spatial distribution patterns. An effective way to circumvent this limitation is to generate an image layout as guidance, which is attempted by a few methods. Nevertheless, these methods fail to generate practically effective layouts due to the diversity of input text and object location. In this paper we push for effective modeling in both text-to-layout generation and layout-to-image synthesis. Specifically, we formulate the text-to-layout generation as a sequence-to-sequence modeling task, and build our model upon Transformer to learn the spatial relationships between objects by modeling the sequential dependencies between them. In the stage of layout-to-image synthesis, we focus on learning the textual-visual semantic alignment per object in the layout to precisely incorporate the input text into the layout-to-image synthesizing process. To evaluate the quality of generated layout, we design a new metric specifically, dubbed Layout Quality Score, which considers both the absolute distribution errors of bounding boxes in the layout and the mutual spatial relationships between them. Extensive experiments on three datasets demonstrate the superior performance of our method over state-of-the-art methods on both predicting the layout and synthesizing the image from the given text.

Dianwen Mei, Wei Zhuo, Jiandong Tian et al. · tencent-ai

Despite the remarkable success of existing methods for few-shot segmentation, there remain two crucial challenges. First, the feature learning for novel classes is suppressed during the training on base classes in that the novel classes are always treated as background. Thus, the semantics of novel classes are not well learned. Second, most of existing methods fail to consider the underlying semantic gap between the support and the query resulting from the representative bias by the scarce support samples. To circumvent these two challenges, we propose to activate the discriminability of novel classes explicitly in both the feature encoding stage and the prediction stage for segmentation. In the feature encoding stage, we design the Semantic-Preserving Feature Learning module (SPFL) to first exploit and then retain the latent semantics contained in the whole input image, especially those in the background that belong to novel classes. In the prediction stage for segmentation, we learn an Self-Refined Online Foreground-Background classifier (SROFB), which is able to refine itself using the high-confidence pixels of query image to facilitate its adaptation to the query image and bridge the support-query semantic gap. Extensive experiments on PASCAL-5$^i$ and COCO-20$^i$ datasets demonstrates the advantages of these two novel designs both quantitatively and qualitatively.

Jing Li, Di Kang, Wenjie Pei et al.

People may perform diverse gestures affected by various mental and physical factors when speaking the same sentences. This inherent one-to-many relationship makes co-speech gesture generation from audio particularly challenging. Conventional CNNs/RNNs assume one-to-one mapping, and thus tend to predict the average of all possible target motions, easily resulting in plain/boring motions during inference. So we propose to explicitly model the one-to-many audio-to-motion mapping by splitting the cross-modal latent code into shared code and motion-specific code. The shared code is expected to be responsible for the motion component that is more correlated to the audio while the motion-specific code is expected to capture diverse motion information that is more independent of the audio. However, splitting the latent code into two parts poses extra training difficulties. Several crucial training losses/strategies, including relaxed motion loss, bicycle constraint, and diversity loss, are designed to better train the VAE. Experiments on both 3D and 2D motion datasets verify that our method generates more realistic and diverse motions than previous state-of-the-art methods, quantitatively and qualitatively. Besides, our formulation is compatible with discrete cosine transformation (DCT) modeling and other popular backbones (\textit{i.e.} RNN, Transformer). As for motion losses and quantitative motion evaluation, we find structured losses/metrics (\textit{e.g.} STFT) that consider temporal and/or spatial context complement the most commonly used point-wise losses (\textit{e.g.} PCK), resulting in better motion dynamics and more nuanced motion details. Finally, we demonstrate that our method can be readily used to generate motion sequences with user-specified motion clips on the timeline.

Wenjie Pei, Shuang Wu, Dianwen Mei et al.

While fine-tuning based methods for few-shot object detection have achieved remarkable progress, a crucial challenge that has not been addressed well is the potential class-specific overfitting on base classes and sample-specific overfitting on novel classes. In this work we design a novel knowledge distillation framework to guide the learning of the object detector and thereby restrain the overfitting in both the pre-training stage on base classes and fine-tuning stage on novel classes. To be specific, we first present a novel Position-Aware Bag-of-Visual-Words model for learning a representative bag of visual words (BoVW) from a limited size of image set, which is used to encode general images based on the similarities between the learned visual words and an image. Then we perform knowledge distillation based on the fact that an image should have consistent BoVW representations in two different feature spaces. To this end, we pre-learn a feature space independently from the object detection, and encode images using BoVW in this space. The obtained BoVW representation for an image can be considered as distilled knowledge to guide the learning of object detector: the extracted features by the object detector for the same image are expected to derive the consistent BoVW representations with the distilled knowledge. Extensive experiments validate the effectiveness of our method and demonstrate the superiority over other state-of-the-art methods.

Yi Li, Wenjie Pei, Zhenyu He

The traditional homography estimation pipeline consists of four main steps: feature detection, feature matching, outlier removal and transformation estimation. Recent deep learning models intend to address the homography estimation problem using a single convolutional network. While these models are trained in an end-to-end fashion to simplify the homography estimation problem, they lack the feature matching step and/or the outlier removal step, which are important steps in the traditional homography estimation pipeline. In this paper, we attempt to build a deep learning model that mimics all four steps in the traditional homography estimation pipeline. In particular, the feature matching step is implemented using the cost volume technique. To remove outliers in the cost volume, we treat this outlier removal problem as a denoising problem and propose a novel self-supervised loss to solve the problem. Extensive experiments on synthetic and real datasets demonstrate that the proposed model outperforms existing deep learning models.

Yufei Xu, Fanxu Meng, Fan Jiang et al.

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical token for each query using a lightweight indexer, and then computing attention only over the selected subset. While the downstream sparse attention scales efficiently, the indexer still scans the entire prefix for every query, introducing an O($L^2$) per-layer bottleneck that becomes prohibitive as context length grows. We propose HISA (Hierarchical Indexed Sparse Attention), a drop-in replacement for the indexer that transforms the search process from a flat token scan into a two-stage hierarchical procedure. First, a block-level coarse filter scores pooled block representatives to prune irrelevant regions. Then, a token-level refinement applies the original indexer only within the remaining candidate blocks. HISA preserves the exact token-level top-k sparsity pattern required by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves a 2$\times$ speedup at 32K context length and 4$\times$ at 128K. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 with HISA, without any fine-tuning. HISA closely matches the original DSA in quality while significantly outperforming block-sparse baselines. Moreover, the token selection sets produced by HISA and the original DSA exhibit a mean IoU greater than 99%, indicating that the efficiency gains come with virtually no impact on selection fidelity.

Fengjun Li, Xin Feng, Fanglin Chen et al.

Typical methods for blind image super-resolution (SR) focus on dealing with unknown degradations by directly estimating them or learning the degradation representations in a latent space. A potential limitation of these methods is that they assume the unknown degradations can be simulated by the integration of various handcrafted degradations (e.g., bicubic downsampling), which is not necessarily true. The real-world degradations can be beyond the simulation scope by the handcrafted degradations, which are referred to as novel degradations. In this work, we propose to learn a latent representation space for degradations, which can be generalized from handcrafted (base) degradations to novel degradations. The obtained representations for a novel degradation in this latent space are then leveraged to generate degraded images consistent with the novel degradation to compose paired training data for SR model. Furthermore, we perform variational inference to match the posterior of degradations in latent representation space with a prior distribution (e.g., Gaussian distribution). Consequently, we are able to sample more high-quality representations for a novel degradation to augment the training data for SR model. We conduct extensive experiments on both synthetic and real-world datasets to validate the effectiveness and advantages of our method for blind super-resolution with novel degradations.

Jingjing Wu, Zhengyao Fang, Pengyuan Lyu et al.

Transcription-only Supervised Text Spotting aims to learn text spotters relying only on transcriptions but no text boundaries for supervision, thus eliminating expensive boundary annotation. The crux of this task lies in locating each transcription in scene text images without location annotations. In this work, we formulate this challenging problem as a Weakly Supervised Cross-modality Contrastive Learning problem, and design a simple yet effective model dubbed WeCromCL that is able to detect each transcription in a scene image in a weakly supervised manner. Unlike typical methods for cross-modality contrastive learning that focus on modeling the holistic semantic correlation between an entire image and a text description, our WeCromCL conducts atomistic contrastive learning to model the character-wise appearance consistency between a text transcription and its correlated region in a scene image to detect an anchor point for the transcription in a weakly supervised manner. The detected anchor points by WeCromCL are further used as pseudo location labels to guide the learning of text spotting. Extensive experiments on four challenging benchmarks demonstrate the superior performance of our model over other methods. Code will be released.

Jingjing Wu, Pengyuan Lyu, Guangming Lu et al.

Typical text spotters follow the two-stage spotting paradigm which detects the boundary for a text instance first and then performs text recognition within the detected regions. Despite the remarkable progress of such spotting paradigm, an important limitation is that the performance of text recognition depends heavily on the precision of text detection, resulting in the potential error propagation from detection to recognition. In this work, we propose the single shot Self-Reliant Scene Text Spotter v2 (SRSTS v2), which circumvents this limitation by decoupling recognition from detection while optimizing two tasks collaboratively. Specifically, our SRSTS v2 samples representative feature points around each potential text instance, and conducts both text detection and recognition in parallel guided by these sampled points. Thus, the text recognition is no longer dependent on detection, thereby alleviating the error propagation from detection to recognition. Moreover, the sampling module is learned under the supervision from both detection and recognition, which allows for the collaborative optimization and mutual enhancement between two tasks. Benefiting from such sampling-driven concurrent spotting framework, our approach is able to recognize the text instances correctly even if the precise text boundaries are challenging to detect. Extensive experiments on four benchmarks demonstrate that our method compares favorably to state-of-the-art spotters.

Zikun Zhou, Kaige Mao, Wenjie Pei et al.

This paper aims to solve the video object segmentation (VOS) task in a scribble-supervised manner, in which VOS models are not only trained by the sparse scribble annotations but also initialized with the sparse target scribbles for inference. Thus, the annotation burdens for both training and initialization can be substantially lightened. The difficulties of scribble-supervised VOS lie in two aspects. On the one hand, it requires the powerful ability to learn from the sparse scribble annotations during training. On the other hand, it demands strong reasoning capability during inference given only a sparse initial target scribble. In this work, we propose a Reliability-Hierarchical Memory Network (RHMNet) to predict the target mask in a step-wise expanding strategy w.r.t. the memory reliability level. To be specific, RHMNet first only uses the memory in the high-reliability level to locate the region with high reliability belonging to the target, which is highly similar to the initial target scribble. Then it expands the located high-reliability region to the entire target conditioned on the region itself and the memories in all reliability levels. Besides, we propose a scribble-supervised learning mechanism to facilitate the learning of our model to predict dense results. It mines the pixel-level relation within the single frame and the frame-level relation within the sequence to take full advantage of the scribble annotations in sequence training samples. The favorable performance on two popular benchmarks demonstrates that our method is promising.

Junwei Wen, Deshui Miao, Guangming Lu et al.

Video Reasoning Segmentation (VRS) aims to segment target objects in videos based on implicit instructions that convey human intent and temporal logic. Existing MLLM-based methods predict masks with a [SEG] token after selecting frames via simple sampling or an auxiliary MLLM, where limited supervision and frame-language similarity rules often yield narrow-scope keyframe choices that weaken holistic temporal understanding and lead to brittle localization in complex multi-object scenes. To address these issues, we introduce RCoT-Seg, a video-of-thought framework that factorizes VRS into temporal video reasoning (TVR) and keyframe target perception (KTP), explicitly separating temporal reasoning from spatial perception. Specifically, in the TVR stage, an agentic keyframe selection module, initialized with a curated CoT-start corpus and refined by GRPO under task-aligned rewards, is proposed to generate and reselect the keyframe through self-evaluation, strengthening moment localization and temporal reasoning. In the KTP stage, RCoT-Seg performs high-resolution segmentation on the selected frame and propagates masks with SAM2-based methods across the sequence, replacing heuristic sampling and external selectors while improving spatial precision and inter-frame consistency. Extensive experimental results demonstrate that the proposed RCoT-Seg achieves favorable performance against the state-of-the-art methods. The code and models will be publicly released at https://github.com/Victor-wjw/RCoT-Seg.

Pingzhi Tang, Ruijie Zhou, Fanxu Meng et al.

Current quantization methods for LLMs predominantly rely on block-wise structures to maintain efficiency, often at the cost of representational flexibility. In this work, we demonstrate that element-wise quantization can be made as efficient as block-wise scaling while providing strictly superior expressive power by modeling the scaling manifold as continuous low-rank matrices ($S = BA$). We propose Low-Rank Decomposed Scaling (LoRDS), a unified framework that rethinks quantization granularity through this low-rank decomposition. By "breaking the blocks" of spatial constraints, LoRDS establishes a seamless efficiency lifecycle: it provides high-fidelity PTQ initialization refined via iterative optimization, enables joint QAT of weights and scaling factors, and facilitates high-rank multiplicative PEFT adaptation. Unlike additive PEFT approaches such as QLoRA, LoRDS enables high-rank weight updates within a low-rank budget while incurring no additional inference overhead. Supported by highly optimized Triton kernels, LoRDS consistently outperforms state-of-the-art baselines across various model families in both quantization and downstream fine-tuning tasks. Notably, on Llama3-8B, our method achieves up to a 27.0% accuracy improvement at 3 bits over NormalFloat quantization and delivers a 1.5x inference speedup on NVIDIA RTX 4090 while enhancing PEFT performance by 9.6% on downstream tasks over 4bit QLoRA, offering a robust and integrated solution for unified compression and adaptation of LLMs.

Jiatong Zhang, Zengwei Yao, Fanglin Chen et al.

Vision Transformers have achieved remarkable progresses, among which Swin Transformer has demonstrated the tremendous potential of Transformer for vision tasks. It surmounts the key challenge of high computational complexity by performing local self-attention within shifted windows. In this work we propose the Semantic-Aware Local-Global Vision Transformer (SALG), to further investigate two potential improvements towards Swin Transformer. First, unlike Swin Transformer that performs uniform partition to produce equal size of regular windows for local self-attention, our SALG performs semantic segmentation in an unsupervised way to explore the underlying semantic priors in the image. As a result, each segmented region can correspond to a semantically meaningful part in the image, potentially leading to more effective features within each of segmented regions. Second, instead of only performing local self-attention within local windows as Swin Transformer does, the proposed SALG performs both 1) local intra-region self-attention for learning fine-grained features within each region and 2) global inter-region feature propagation for modeling global dependencies among all regions. Consequently, our model is able to obtain the global view when learning features for each token, which is the essential advantage of Transformer. Owing to the explicit modeling of the semantic priors and the proposed local-global modeling mechanism, our SALG is particularly advantageous for small-scale models when the modeling capacity is not sufficient for other models to learn semantics implicitly. Extensive experiments across various vision tasks demonstrates the merit of our model over other vision Transformers, especially in the small-scale modeling scenarios.

Jiapeng Su, Qi Fan, Guangming Lu et al.

Few-shot semantic segmentation (FSS) has achieved great success on segmenting objects of novel classes, supported by only a few annotated samples. However, existing FSS methods often underperform in the presence of domain shifts, especially when encountering new domain styles that are unseen during training. It is suboptimal to directly adapt or generalize the entire model to new domains in the few-shot scenario. Instead, our key idea is to adapt a small adapter for rectifying diverse target domain styles to the source domain. Consequently, the rectified target domain features can fittingly benefit from the well-optimized source domain segmentation model, which is intently trained on sufficient source domain data. Training domain-rectifying adapter requires sufficiently diverse target domains. We thus propose a novel local-global style perturbation method to simulate diverse potential target domains by perturbating the feature channel statistics of the individual images and collective statistics of the entire source domain, respectively. Additionally, we propose a cyclic domain alignment module to facilitate the adapter effectively rectifying domains using a reverse domain rectification supervision. The adapter is trained to rectify the image features from diverse synthesized target domains to align with the source domain. During testing on target domains, we start by rectifying the image features and then conduct few-shot segmentation on the domain-rectified features. Extensive experiments demonstrate the effectiveness of our method, achieving promising results on cross-domain few-shot semantic segmentation tasks. Our code is available at https://github.com/Matt-Su/DR-Adapter.

Wenjie Pei, Tongqi Xia, Fanglin Chen et al.

As a prominent parameter-efficient fine-tuning technique in NLP, prompt tuning is being explored its potential in computer vision. Typical methods for visual prompt tuning follow the sequential modeling paradigm stemming from NLP, which represents an input image as a flattened sequence of token embeddings and then learns a set of unordered parameterized tokens prefixed to the sequence representation as the visual prompts for task adaptation of large vision models. While such sequential modeling paradigm of visual prompt has shown great promise, there are two potential limitations. First, the learned visual prompts cannot model the underlying spatial relations in the input image, which is crucial for image encoding. Second, since all prompt tokens play the same role of prompting for all image tokens without distinction, it lacks the fine-grained prompting capability, i.e., individual prompting for different image tokens. In this work, we propose the \mymodel model (\emph{SA$^2$VP}), which learns a two-dimensional prompt token map with equal (or scaled) size to the image token map, thereby being able to spatially align with the image map. Each prompt token is designated to prompt knowledge only for the spatially corresponding image tokens. As a result, our model can conduct individual prompting for different image tokens in a fine-grained manner. Moreover, benefiting from the capability of preserving the spatial structure by the learned prompt token map, our \emph{SA$^2$VP} is able to model the spatial relations in the input image, leading to more effective prompting. Extensive experiments on three challenging benchmarks for image classification demonstrate the superiority of our model over other state-of-the-art methods for visual prompt tuning. Code is available at \emph{https://github.com/tommy-xq/SA2VP}.

Zhenhua Ning, Xin Li, Jun Yu et al.

While 3D Gaussian Splatting (3DGS) has demonstrated impressive real-time rendering performance, its efficacy remains constrained by a reliance on heuristic density control. Despite numerous refinements to these handcrafted rules, such methods inherently lack the flexibility to adapt to diverse scenes with complex geometries. In this paper, we propose a paradigm shift for density control from rigid heuristics to fully learnable policies. Specifically, we introduce \textbf{LeGS}, a framework that reformulates density control as a parameterized policy network optimized via Reinforcement Learning (RL). Central to our approach is the tailored effective reward function grounded in sensitivity analysis, which precisely quantifies the marginal contribution of individual Gaussians to reconstruction quality. To maintain computational tractability, we derive a closed-form solution that reduces the complexity of reward calculation from $O(N^2)$ to $O(N)$. Extensive experiments on the Mip-NeRF 360, Tanks \& Temples, and Deep Blending datasets demonstrate that \textbf{LeGS} significantly outperforms state-of-the-art methods, striking a superior balance between reconstruction quality and efficiency. The code will be released at https://github.com/AaronNZH/LeGS

Qizhong Tan, Zhuotao Tian, Guangming Lu et al.

Existing Video Large Language Models (Video LLMs) struggle with complex video understanding, exhibiting limited reasoning capabilities and potential hallucinations. In particular, these methods tend to perform reasoning solely relying on the pretrained inherent reasoning rationales whilst lacking perception-aware adaptation to the input video content. To address this, we propose \textbf{Video-ToC}, a novel video reasoning framework that enhances video understanding through tree-of-cue reasoning. Specifically, our approach introduces three key innovations: (1) A tree-guided visual cue localization mechanism, which endows the model with enhanced fine-grained perceptual capabilities through structured reasoning patterns; (2) A reasoning-demand reward mechanism, which dynamically adjusts the reward value for reinforcement learning (RL) based on the estimation of reasoning demands, enabling on-demand incentives for more effective reasoning strategies; and (3) An automated annotation pipeline that constructs the Video-ToC-SFT-1k and Video-ToC-RL-2k datasets for supervised fine-tuning (SFT) and RL training, respectively. Extensive evaluations on six video understanding benchmarks and a video hallucination benchmark demonstrate the superiority of Video-ToC over baselines and recent methods. Code is available at https://github.com/qizhongtan/Video-ToC.

Zhengyao Fang, Pengyuan Lyu, Jingjing Wu et al.

Scene text editing aims to modify text content within scene images while maintaining style consistency. Traditional methods achieve this by explicitly disentangling style and content from the source image and then fusing the style with the target content, while ensuring content consistency using a pre-trained recognition model. Despite notable progress, these methods suffer from complex pipelines, leading to suboptimal performance in complex scenarios. In this work, we introduce Recognition-Synergistic Scene Text Editing (RS-STE), a novel approach that fully exploits the intrinsic synergy of text recognition for editing. Our model seamlessly integrates text recognition with text editing within a unified framework, and leverages the recognition model's ability to implicitly disentangle style and content while ensuring content consistency. Specifically, our approach employs a multi-modal parallel decoder based on transformer architecture, which predicts both text content and stylized images in parallel. Additionally, our cyclic self-supervised fine-tuning strategy enables effective training on unpaired real-world data without ground truth, enhancing style and content consistency through a twice-cyclic generation process. Built on a relatively simple architecture, RS-STE achieves state-of-the-art performance on both synthetic and real-world benchmarks, and further demonstrates the effectiveness of leveraging the generated hard cases to boost the performance of downstream recognition tasks. Code is available at https://github.com/ZhengyaoFang/RS-STE.

Jiahuan Wang, Yuxin Chen, Jun Yu et al.

Adapting pretrained diffusion-based generative models for text-driven image editing with negligible tuning overhead has demonstrated remarkable potential. A classical adaptation paradigm, as followed by these methods, first infers the generative trajectory inversely for a given source image by image inversion, then performs image editing along the inferred trajectory guided by the target text prompts. However, the performance of image editing is heavily limited by the approximation errors introduced during image inversion by diffusion models, which arise from the absence of exact supervision in the intermediate generative steps. To circumvent this issue, we investigate the parameter-efficient adaptation of binary-quantized generative models for image editing, and leverage their inherent characteristic that the exact intermediate quantized representations of a source image are attainable, enabling more effective supervision for precise image inversion. Specifically, we propose EditInfinity, which adapts \emph{Infinity}, a binary-quantized generative model, for image editing. We propose an efficient yet effective image inversion mechanism that integrates text prompting rectification and image style preservation, enabling precise image inversion. Furthermore, we devise a holistic smoothing strategy which allows our EditInfinity to perform image editing with high fidelity to source images and precise semantic alignment to the text prompts. Extensive experiments on the PIE-Bench benchmark across `add', `change', and `delete' editing operations, demonstrate the superior performance of our model compared to state-of-the-art diffusion-based baselines. Code available at: https://github.com/yx-chen-ust/EditInfinity.

Yushuai Sun, Zikun Zhou, Dongmei Jiang et al.

Asymmetric retrieval is a typical scenario in real-world retrieval systems, where compatible models of varying capacities are deployed on platforms with different resource configurations. Existing methods generally train pre-defined networks or subnetworks with capacities specifically designed for pre-determined platforms, using compatible learning. Nevertheless, these methods suffer from limited flexibility for multi-platform deployment. For example, when introducing a new platform into the retrieval systems, developers have to train an additional model at an appropriate capacity that is compatible with existing models via backward-compatible learning. In this paper, we propose a Prunable Network with self-compatibility, which allows developers to generate compatible subnetworks at any desired capacity through post-training pruning. Thus it allows the creation of a sparse subnetwork matching the resources of the new platform without additional training. Specifically, we optimize both the architecture and weight of subnetworks at different capacities within a dense network in compatible learning. We also design a conflict-aware gradient integration scheme to handle the gradient conflicts between the dense network and subnetworks during compatible learning. Extensive experiments on diverse benchmarks and visual backbones demonstrate the effectiveness of our method. Our code and model are available at https://github.com/Bunny-Black/PrunNet.

Ruijie Zhou, Fanxu Meng, Yufei Xu et al.

DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.

Wenjie Pei, Weina Xu, Zongze Wu et al.

The crux of graph classification lies in the effective representation learning for the entire graph. Typical graph neural networks focus on modeling the local dependencies when aggregating features of neighboring nodes, and obtain the representation for the entire graph by aggregating node features. Such methods have two potential limitations: 1) the global node saliency w.r.t. graph classification is not explicitly modeled, which is crucial since different nodes may have different semantic relevance to graph classification; 2) the graph representation directly aggregated from node features may have limited effectiveness to reflect graph-level information. In this work, we propose the Saliency-Aware Regularized Graph Neural Network (SAR-GNN) for graph classification, which consists of two core modules: 1) a traditional graph neural network serving as the backbone for learning node features and 2) the Graph Neural Memory designed to distill a compact graph representation from node features of the backbone. We first estimate the global node saliency by measuring the semantic similarity between the compact graph representation and node features. Then the learned saliency distribution is leveraged to regularize the neighborhood aggregation of the backbone, which facilitates the message passing of features for salient nodes and suppresses the less relevant nodes. Thus, our model can learn more effective graph representation. We demonstrate the merits of SAR-GNN by extensive experiments on seven datasets across various types of graph data. Code will be released.

Jingwen Zhang, Zikun Zhou, Guangming Lu et al.

3D single object tracking remains a challenging problem due to the sparsity and incompleteness of the point clouds. Existing algorithms attempt to address the challenges in two strategies. The first strategy is to learn dense geometric features based on the captured sparse point cloud. Nevertheless, it is quite a formidable task since the learned dense geometric features are with high uncertainty for depicting the shape of the target object. The other strategy is to aggregate the sparse geometric features of multiple templates to enrich the shape information, which is a routine solution in 2D tracking. However, aggregating the coarse shape representations can hardly yield a precise shape representation. Different from 2D pixels, 3D points of different frames can be directly fused by coordinate transform, i.e., shape completion. Considering that, we propose to construct a synthetic target representation composed of dense and complete point clouds depicting the target shape precisely by shape completion for robust 3D tracking. Specifically, we design a voxelized 3D tracking framework with shape completion, in which we propose a quality-aware shape completion mechanism to alleviate the adverse effect of noisy historical predictions. It enables us to effectively construct and leverage the synthetic target representation. Besides, we also develop a voxelized relation modeling module and box refinement module to improve tracking performance. Favorable performance against state-of-the-art algorithms on three benchmarks demonstrates the effectiveness and generalization ability of our method.

Bo Zhang, Zhuotao Tian, Xin Tao et al.

Amodal segmentation is a challenging task that aims to predict the complete geometric shape of objects, including their occluded regions. Although existing methods primarily focus on amodal segmentation within the training domain, these approaches often lack the generalization capacity to extend effectively to novel object categories and unseen contexts. This paper introduces Amodal SAM, a unified framework that leverages SAM (Segment Anything Model) for both amodal image and amodal video segmentation. Amodal SAM preserves the powerful generalization ability of SAM while extending its inherent capabilities to the amodal segmentation task. The improvements lie in three aspects: (1) a lightweight Spatial Completion Adapter that enables occluded region reconstruction, (2) a Target-Aware Occlusion Synthesis (TAOS) pipeline that addresses the scarcity of amodal annotations by generating diverse synthetic training data, and (3) novel learning objectives that enforce regional consistency and topological regularization. Extensive experiments demonstrate that Amodal SAM achieves state-of-the-art performance on standard benchmarks, while simultaneously exhibiting robust generalization to novel scenarios. We anticipate that this research will advance the field toward practical amodal segmentation systems capable of operating effectively in unconstrained real-world environments.

Kuiyuan Zhang, Wenjie Pei, Rushi Lan et al.

Deepfakes are AI-synthesized multimedia data that may be abused for spreading misinformation. Deepfake generation involves both visual and audio manipulation. To detect audio-visual deepfakes, previous studies commonly employ two relatively independent sub-models to learn audio and visual features, respectively, and fuse them subsequently for deepfake detection. However, this may underutilize the inherent correlations between audio and visual features. Moreover, utilizing two isolated feature learning sub-models can result in redundant neural layers, making the overall model inefficient and impractical for resource-constrained environments. In this work, we design a lightweight network for audio-visual deepfake detection via a single-stream multi-modal learning framework. Specifically, we introduce a collaborative audio-visual learning block to efficiently integrate multi-modal information while learning the visual and audio features. By iteratively employing this block, our single-stream network achieves a continuous fusion of multi-modal features across its layers. Thus, our network efficiently captures visual and audio features without the need for excessive block stacking, resulting in a lightweight network design. Furthermore, we propose a multi-modal classification module that can boost the dependence of the visual and audio classifiers on modality content. It also enhances the whole resistance of the video classifier against the mismatches between audio and visual modalities. We conduct experiments on the DF-TIMIT, FakeAVCeleb, and DFDC benchmark datasets. Compared to state-of-the-art audio-visual joint detection methods, our method is significantly lightweight with only 0.48M parameters, yet it achieves superiority in both uni-modal and multi-modal deepfakes, as well as in unseen types of deepfakes.

Zikun Zhou, Yushuai Sun, Wenjie Pei et al.

The traditional paradigm to update retrieval models requires re-computing the embeddings of the gallery data, a time-consuming and computationally intensive process known as backfilling. To circumvent backfilling, Backward-Compatible Learning (BCL) has been widely explored, which aims to train a new model compatible with the old one. Many previous works focus on effectively aligning the embeddings of the new model with those of the old one to enhance the backward-compatibility. Nevertheless, such strong alignment constraints would compromise the discriminative ability of the new model, particularly when different classes are closely clustered and hard to distinguish in the old feature space. To address this issue, we propose to relax the constraints by introducing perturbations to the old feature prototypes. This allows us to align the new feature space with a pseudo-old feature space defined by these perturbed prototypes, thereby preserving the discriminative ability of the new model in backward-compatible learning. We have developed two approaches for calculating the perturbations: Neighbor-Driven Prototype Perturbation (NDPP) and Optimization-Driven Prototype Perturbation (ODPP). Particularly, they take into account the feature distributions of not only the old but also the new models to obtain proper perturbations along with new model updating. Extensive experiments on the landmark and commodity datasets demonstrate that our approaches perform favorably against state-of-the-art BCL algorithms.

Yanchen Liu, Yanan Sun, Zhening Xing et al.

Existing text-to-video methods struggle to transfer motion smoothly from a reference object to a target object with significant differences in appearance or structure between them. To address this challenge, we introduce MotionShot, a training-free framework capable of parsing reference-target correspondences in a fine-grained manner, thereby achieving high-fidelity motion transfer while preserving coherence in appearance. To be specific, MotionShot first performs semantic feature matching to ensure high-level alignments between the reference and target objects. It then further establishes low-level morphological alignments through reference-to-target shape retargeting. By encoding motion with temporal attention, our MotionShot can coherently transfer motion across objects, even in the presence of significant appearance and structure disparities, demonstrated by extensive experiments. The project page is available at: https://motionshot.github.io/.

Zhenhua Ning, Zhuotao Tian, Shaoshuai Shi et al.

Recent advances in point cloud perception have demonstrated remarkable progress in scene understanding through vision-language alignment leveraging large language models (LLMs). However, existing methods may still encounter challenges in handling complex instructions that require accurate spatial reasoning, even if the 3D point cloud data provides detailed spatial cues such as size and position for identifying the targets. To tackle this issue, we propose Relevant Reasoning Segmentation (R$^2$S), a reasoning-based segmentation framework. The framework emulates human cognitive processes by decomposing spatial reasoning into two sequential stages: first identifying relevant elements, then processing instructions guided by their associated visual priors. Furthermore, acknowledging the inadequacy of existing datasets in complex reasoning tasks, we introduce 3D ReasonSeg, a reasoning-based segmentation dataset comprising 25,185 training samples and 3,966 validation samples with precise annotations. Both quantitative and qualitative experiments demonstrate that the R$^2$S and 3D ReasonSeg effectively endow 3D point cloud perception with stronger spatial reasoning capabilities, and we hope that they can serve as a new baseline and benchmark for future work.

Yanan Sun, Yanchen Liu, Yinhao Tang et al.

The field of text-to-image (T2I) generation has made significant progress in recent years, largely driven by advancements in diffusion models. Linguistic control enables effective content creation, but struggles with fine-grained control over image generation. This challenge has been explored, to a great extent, by incorporating additional user-supplied spatial conditions, such as depth maps and edge maps, into pre-trained T2I models through extra encoding. However, multi-control image synthesis still faces several challenges. Specifically, current approaches are limited in handling free combinations of diverse input control signals, overlook the complex relationships among multiple spatial conditions, and often fail to maintain semantic alignment with provided textual prompts. This can lead to suboptimal user experiences. To address these challenges, we propose AnyControl, a multi-control image synthesis framework that supports arbitrary combinations of diverse control signals. AnyControl develops a novel Multi-Control Encoder that extracts a unified multi-modal embedding to guide the generation process. This approach enables a holistic understanding of user inputs, and produces high-quality, faithful results under versatile control signals, as demonstrated by extensive quantitative and qualitative evaluations. Our project page is available in https://any-control.github.io.

Zikun Zhou, Jianqiu Chen, Wenjie Pei et al.

The crux of long-term tracking lies in the difficulty of tracking the target with discontinuous moving caused by out-of-view or occlusion. Existing long-term tracking methods follow two typical strategies. The first strategy employs a local tracker to perform smooth tracking and uses another re-detector to detect the target when the target is lost. While it can exploit the temporal context like historical appearances and locations of the target, a potential limitation of such strategy is that the local tracker tends to misidentify a nearby distractor as the target instead of activating the re-detector when the real target is out of view. The other long-term tracking strategy tracks the target in the entire image globally instead of local tracking based on the previous tracking results. Unfortunately, such global tracking strategy cannot leverage the temporal context effectively. In this work, we combine the advantages of both strategies: tracking the target in a global view while exploiting the temporal context. Specifically, we perform global tracking via ensemble of local trackers spreading the full image. The smooth moving of the target can be handled steadily by one local tracker. When the local tracker accidentally loses the target due to suddenly discontinuous moving, another local tracker close to the target is then activated and can readily take over the tracking to locate the target. While the activated local tracker performs tracking locally by leveraging the temporal context, the ensemble of local trackers renders our model the global view for tracking. Extensive experiments on six datasets demonstrate that our method performs favorably against state-of-the-art algorithms.

Jing Xu, Xinglin Pan, Xu Luo et al.

Few-shot classification aims to adapt classifiers to novel classes with a few training samples. However, the insufficiency of training data may cause a biased estimation of feature distribution in a certain class. To alleviate this problem, we present a simple yet effective feature rectification method by exploring the category correlation between novel and base classes as the prior knowledge. We explicitly capture such correlation by mapping features into a latent vector with dimension matching the number of base classes, treating it as the logarithm probability of the feature over base classes. Based on this latent vector, the rectified feature is directly constructed by a decoder, which we expect maintaining category-related information while removing other stochastic factors, and consequently being closer to its class centroid. Furthermore, by changing the temperature value in softmax, we can re-balance the feature rectification and reconstruction for better performance. Our method is generic, flexible and agnostic to any feature extractor and classifier, readily to be embedded into existing FSL approaches. Experiments verify that our method is capable of rectifying biased features, especially when the feature is far from the class centroid. The proposed approach consistently obtains considerable performance gains on three widely used benchmarks, evaluated with different backbones and classifiers. The code will be made public.

Xin Li, Qiao Liu, Wenjie Pei et al.

Along with the rapid progress of visual tracking, existing benchmarks become less informative due to redundancy of samples and weak discrimination between current trackers, making evaluations on all datasets extremely time-consuming. Thus, a small and informative benchmark, which covers all typical challenging scenarios to facilitate assessing the tracker performance, is of great interest. In this work, we develop a principled way to construct a small and informative tracking benchmark (ITB) with 7% out of 1.2 M frames of existing and newly collected datasets, which enables efficient evaluation while ensuring effectiveness. Specifically, we first design a quality assessment mechanism to select the most informative sequences from existing benchmarks taking into account 1) challenging level, 2) discriminative strength, 3) and density of appearance variations. Furthermore, we collect additional sequences to ensure the diversity and balance of tracking scenarios, leading to a total of 20 sequences for each scenario. By analyzing the results of 15 state-of-the-art trackers re-trained on the same data, we determine the effective methods for robust tracking under each scenario and demonstrate new challenges for future research direction in this field.

Haobo Ji, Xin Feng, Wenjie Pei et al.

While Transformer has achieved remarkable performance in various high-level vision tasks, it is still challenging to exploit the full potential of Transformer in image restoration. The crux lies in the limited depth of applying Transformer in the typical encoder-decoder framework for image restoration, resulting from heavy self-attention computation load and inefficient communications across different depth (scales) of layers. In this paper, we present a deep and effective Transformer-based network for image restoration, termed as U2-Former, which is able to employ Transformer as the core operation to perform image restoration in a deep encoding and decoding space. Specifically, it leverages the nested U-shaped structure to facilitate the interactions across different layers with different scales of feature maps. Furthermore, we optimize the computational efficiency for the basic Transformer block by introducing a feature-filtering mechanism to compress the token representation. Apart from the typical supervision ways for image restoration, our U2-Former also performs contrastive learning in multiple aspects to further decouple the noise component from the background image. Extensive experiments on various image restoration tasks, including reflection removal, rain streak removal and dehazing respectively, demonstrate the effectiveness of the proposed U2-Former.

Zebin Lin, Wenjie Pei, Fanglin Chen et al.

Typical methods for pedestrian detection focus on either tackling mutual occlusions between crowded pedestrians, or dealing with the various scales of pedestrians. Detecting pedestrians with substantial appearance diversities such as different pedestrian silhouettes, different viewpoints or different dressing, remains a crucial challenge. Instead of learning each of these diverse pedestrian appearance features individually as most existing methods do, we propose to perform contrastive learning to guide the feature learning in such a way that the semantic distance between pedestrians with different appearances in the learned feature space is minimized to eliminate the appearance diversities, whilst the distance between pedestrians and background is maximized. To facilitate the efficiency and effectiveness of contrastive learning, we construct an exemplar dictionary with representative pedestrian appearances as prior knowledge to construct effective contrastive training pairs and thus guide contrastive learning. Besides, the constructed exemplar dictionary is further leveraged to evaluate the quality of pedestrian proposals during inference by measuring the semantic distance between the proposal and the exemplar dictionary. Extensive experiments on both daytime and nighttime pedestrian detection validate the effectiveness of the proposed method.

Chaozheng Wang, Shuzheng Gao, Cuiyun Gao et al.

Real-world data usually present long-tailed distributions. Training on imbalanced data tends to render neural networks perform well on head classes while much worse on tail classes. The severe sparseness of training instances for the tail classes is the main challenge, which results in biased distribution estimation during training. Plenty of efforts have been devoted to ameliorating the challenge, including data re-sampling and synthesizing new training instances for tail classes. However, no prior research has exploited the transferable knowledge from head classes to tail classes for calibrating the distribution of tail classes. In this paper, we suppose that tail classes can be enriched by similar head classes and propose a novel distribution calibration approach named as label-Aware Distribution Calibration LADC. LADC transfers the statistics from relevant head classes to infer the distribution of tail classes. Sampling from calibrated distribution further facilitates re-balancing the classifier. Experiments on both image and text long-tailed datasets demonstrate that LADC significantly outperforms existing methods.The visualization also shows that LADC provides a more accurate distribution estimation.

Zengwei Yao, Wenjie Pei, Fanglin Chen et al.

The crux of single-channel speech separation is how to encode the mixture of signals into such a latent embedding space that the signals from different speakers can be precisely separated. Existing methods for speech separation either transform the speech signals into frequency domain to perform separation or seek to learn a separable embedding space by constructing a latent domain based on convolutional filters. While the latter type of methods learning an embedding space achieves substantial improvement for speech separation, we argue that the embedding space defined by only one latent domain does not suffice to provide a thoroughly separable encoding space for speech separation. In this paper, we propose the Stepwise-Refining Speech Separation Network (SRSSN), which follows a coarse-to-fine separation framework. It first learns a 1-order latent domain to define an encoding space and thereby performs a rough separation in the coarse phase. Then the proposed SRSSN learns a new latent domain along each basis function of the existing latent domain to obtain a high-order latent domain in the refining phase, which enables our model to perform a refining separation to achieve a more precise speech separation. We demonstrate the effectiveness of our SRSSN by conducting extensive experiments, including speech separation in a clean (noise-free) setting on WSJ0-2/3mix datasets as well as in noisy/reverberant settings on WHAM!/WHAMR! datasets. Furthermore, we also perform experiments of speech recognition on separated speech signals by our model to evaluate the performance of speech separation indirectly.

Xin Feng, Wenjie Pei, Fengjun Li et al.

Most existing methods for image inpainting focus on learning the intra-image priors from the known regions of the current input image to infer the content of the corrupted regions in the same image. While such methods perform well on images with small corrupted regions, it is challenging for these methods to deal with images with large corrupted area due to two potential limitations: 1) such methods tend to overfit each single training pair of images relying solely on the intra-image prior knowledge learned from the limited known area; 2) the inter-image prior knowledge about the general distribution patterns of visual semantics, which can be transferred across images sharing similar semantics, is not exploited. In this paper, we propose the Generative Memory-Guided Semantic Reasoning Model (GM-SRM), which not only learns the intra-image priors from the known regions, but also distills the inter-image reasoning priors to infer the content of the corrupted regions. In particular, the proposed GM-SRM first pre-learns a generative memory from the whole training data to capture the semantic distribution patterns in a global view. Then the learned memory are leveraged to retrieve the matching inter-image priors for the current corrupted image to perform semantic reasoning during image inpainting. While the intra-image priors are used for guaranteeing the pixel-level content consistency, the inter-image priors are favorable for performing high-level semantic reasoning, which is particularly effective for inferring semantic content for large corrupted area. Extensive experiments on Paris Street View, CelebA-HQ, and Places2 benchmarks demonstrate that our GM-SRM outperforms the state-of-the-art methods for image inpainting in terms of both the visual quality and quantitative metrics.

Jing Li, Di Kang, Wenjie Pei et al.

Generating conversational gestures from speech audio is challenging due to the inherent one-to-many mapping between audio and body motions. Conventional CNNs/RNNs assume one-to-one mapping, and thus tend to predict the average of all possible target motions, resulting in plain/boring motions during inference. In order to overcome this problem, we propose a novel conditional variational autoencoder (VAE) that explicitly models one-to-many audio-to-motion mapping by splitting the cross-modal latent code into shared code and motion-specific code. The shared code mainly models the strong correlation between audio and motion (such as the synchronized audio and motion beats), while the motion-specific code captures diverse motion information independent of the audio. However, splitting the latent code into two parts poses training difficulties for the VAE model. A mapping network facilitating random sampling along with other techniques including relaxed motion loss, bicycle constraint, and diversity loss are designed to better train the VAE. Experiments on both 3D and 2D motion datasets verify that our method generates more realistic and diverse motions than state-of-the-art methods, quantitatively and qualitatively. Finally, we demonstrate that our method can be readily used to generate motion sequences with user-specified motion clips on the timeline. Code and more results are at https://jingli513.github.io/audio2gestures.