论文检索

Biao Yang, Muqi Huang, Yuhui Zhang, Yun Xiong, Kun Zhou, Xi Chen, Shiyang Zhou, Huishuai Bao, Chuan Li, Feng Shi 等

Traditional point-based image editing methods rely on iterative latent optimization or geometric transformations, which are either inefficient in their processing or fail to capture the semantic relationships within the image. These methods often overlook the powerful yet underutilized image editing capabilities inherent in pre-trained diffusion models. In this work, we propose a novel one-step point-based image editing method, named \textbf{AttentionDrag}, which leverages the inherent latent knowledge and feature correlations within pre-trained diffusion models for image editing tasks. This framework enables semantic consistency and high-quality manipulation without the need for extensive re-optimization or retraining. Specifically, we reutilize the latent correlations knowledge learned by the self-attention mechanism in the U-Net module during the DDIM inversion process to automatically identify and adjust relevant image regions, ensuring semantic validity and consistency. Additionally, AttentionDrag adaptively generates masks to guide the editing process, enabling precise and context-aware modifications with friendly interaction. Our results demonstrate a performance that surpasses most state-of-the-art methods with significantly faster speeds, showing a more efficient and semantically coherent solution for point-based image editing tasks. Code is released at: https://github.com/GPlaying/AttentionDrag.

Rui Yan, Jin Wang, Hongyu Qu, Xiaoyu Du, Dong Zhang, Jinhui Tang, Tieniu Tan

Recently, adapting Vision Language Models (VLMs) to zero-shot visual classification by tuning class embedding with a few prompts (Test-time Prompt Tuning, TPT) or replacing class names with generated visual samples (support-set) has shown promising results. However, TPT cannot avoid the semantic gap between modalities while the support-set cannot be tuned. To this end, we draw on each other's strengths and propose a novel framework, namely TEst-time Support-set Tuning for zero-shot Video Classification (TEST-V). It first dilates the support-set with multiple prompts (Multi-prompting Support-set Dilation, MSD) and then erodes the support-set via learnable weights to mine key cues dynamically (Temporal-aware Support-set Erosion, TSE). Specifically, i) MSD expands the support samples for each class based on multiple prompts inquired from LLMs to enrich the diversity of the support-set. ii) TSE tunes the support-set with factorized learnable weights according to the temporal prediction consistency in a self-supervised manner to dig pivotal supporting cues for each class. TEST-V achieves state-of-the-art results across four benchmarks and shows good interpretability.

Lina Wei, Yuhang Ma, Zhongsheng Lin, Fangfang Wang, Canghong Jin, Hanbin Zhao, Dapeng Chen

Multimodal perception, which integrates vision and touch, is increasingly demonstrating its significance in domains such as embodied intelligence and human-computer interaction. However, in open-world scenarios, multimodal data streams face significant challenges, including catastrophic forgetting and overfitting, during few-shot class incremental learning (FSCIL), leading to a severe degradation in model performance. In this work, we propose a novel approach named Few-Shot Incremental Multi-modal Learning via Touch Guidance and Imaginary Vision Synthesis (TIFS). Our method leverages vision imagination synthesis to enhance the semantic understanding and integrates touch and vision fusion to improve the problem of modal imbalance. Specifically, we introduce a framework that employs touch-guided vision information for cross-modal contrastive learning to address the challenges of few-shot learning. Additionally, we incorporate multiple learning mechanisms, including regularization, memory mechanisms, and attention mechanisms, to mitigate catastrophic forgetting during multi-incremental step learning. Experimental results on the Touch and Go and VisGel datasets demonstrate that the TIFS framework exhibits robust continuous learning capabilities and strong generalization performance in touch-vision few-shot incremental learning tasks. Our code is available at https://github.com/Vision-Multimodal-Lab-HZCU/TIFS.

Yongqi Wang, Xinxiao Wu, Shuo Yang

Open-vocabulary video visual relationship detection aims to detect objects and their relationships in videos without being restricted by predefined object or relationship categories. Existing methods leverage the rich semantic knowledge of pre-trained vision-language models such as CLIP to identify novel categories. They typically adopt a cascaded pipeline to first detect objects and then classify relationships based on the detected objects, which may lead to error propagation and thus suboptimal performance. In this paper, we propose Mutual EnhancemenT of Objects and Relationships (METOR), a query-based unified framework to jointly model and mutually enhance object detection and relationship classification in open-vocabulary scenarios. Under this framework, we first design a CLIP-based contextual refinement encoding module that extracts visual contexts of objects and relationships to refine the encoding of text features and object queries, thus improving the generalization of encoding to novel categories. Then we propose an iterative enhancement module to alternatively enhance the representations of objects and relationships by fully exploiting their interdependence to improve recognition performance. Extensive experiments on two public datasets, VidVRD and VidOR, demonstrate that our framework achieves state-of-the-art performance. Codes are at https://github.com/wangyongqi558/METOR.

Xiaoxi Sun, Jianxin Liang, Yueqian Wang, Huishuai Zhang, Dongyan Zhao

Understanding small visual objects is crucial in fields such as video surveillance, remote sensing, and autonomous driving. In this paper, we investigate the capability of advanced large vision-language models (LVLMs) to recognize and interpret small objects in visual data. To this end, we curate a specialized dataset for evaluating fine-grained visual hallucinations, incorporating two object categories and three types of hallucinations. First, we assess 11 state-of-the-art LVLMs, yielding several key insights, as anticipated, LVLMs perform significantly worse on queries related to small objects compared to regular-sized ones, with performance on regular objects proving to be an unreliable predictor of that on small objects. This finding underscores the need for dedicated research on fine-grained visual hallucinations. Second, we evaluate three training-free methods: Scaffold, Chain of Thought (CoT), and Image Resizing, all of which result in varying degrees of improvement. Furthermore, we conduct a series of detailed ablation studies on the visual encoders of Eagle-X5, examining their performance across fine-grained visual hallucination tasks. Our findings reveal that ConvNeXt architecture is critical for object existence recognition tasks. In contrast, for mitigating other types of hallucinations, integrating information from multiple visual encoders is significantly more effective than relying on a single encoder. These results highlight several promising directions for advancing small object recognition with LVLMs.

Xuesong Nie, Haoyuan Jin, Vijayakumar Bhagavatula, Xiaofeng Liu

Predictive modeling of unannotated spatiotemporal data presents inherent challenges, primarily due to the highly entangled visual dynamics in real-world scenes. To tackle these complexities, we introduce a novel insight through Disentangling Deterministic and Probabilistic (DDP) modeling. We note a key observation in spatiotemporal data where low-level details typically remain stable, whereas high-level motion frequently exhibits dynamic variations. The core motivation involves constructing two distinct pathways in the latent space: a deterministic path and a probabilistic path. The probabilistic path begins by defining the motion flow, which explicitly describes complex many-to-many motion patterns between patches, and models its probabilistic distribution using a motion diffuser. The deterministic path incorporates a spectral-aware enhancer to retain and amplify visual details in the frequency domain. These designs ensure visual consistency while also capturing intricate long-term motion dynamics. Extensive experiments demonstrate the superiority of DDP across diverse scenario evaluations.

Yujian Liu, Shidang Xu, Jing Guo, Dingbin Wang, Zairan Wang, Xianfeng Tan, Xiaoli Liu

Generating talking avatar driven by audio remains a significant challenge. Existing methods typically require high computational costs and often lack sufficient facial detail and realism, making them unsuitable for applications that demand high real-time performance and visual quality. Additionally, while some methods can synchronize lip movement, they still face issues with consistency between facial expressions and upper body movement, particularly during silent periods. In this paper, we introduce SyncAnimation, the first NeRF-based method that achieves audio-driven, stable, and real-time generation of speaking avatar by combining generalized audio-to-pose matching and audio-to-expression synchronization. By integrating AudioPose Syncer and AudioEmotion Syncer, SyncAnimation achieves high-precision poses and expression generation, progressively producing audio-synchronized upper body, head, and lip shapes. Furthermore, the High-Synchronization Human Renderer ensures seamless integration of the head and upper body, and achieves audio-sync lip. The project page can be found at https://syncanimation.github.io.

Yingge Liu, Dawei Dai, Xiangling Hou, Shilin Zhao, Guoyin Wang

In contrast with human sketching, which pre-conceptualizes outlines and features, conventional sketch retrieval models rely primarily rely on pixel-level processing and feature extraction, limiting their ability to capture early sketch intent. Consequently, these models are susceptible to subjective stroke noise, reducing retrieval accuracy. To address this issue, we propose a novel on-the-fly noise stroke retrieval framework designed to align with human sketch-drawing cognition. The proposed framework introduces two core innovations. (i) A stroke consistency detection module that effectively discriminates and suppresses noise strokes by quantifying the structural similarity between the current stroke and the target image, as well as its alignment with key skeletal components. (ii) An adaptive gated mixture of experts module that dynamically selects and integrates features from multiple expert networks during the early, sparse stages of sketching, thereby capturing relevant information with greater precision. Experimental results across diverse sketch datasets demonstrate that the proposed method effectively identifies and suppresses early noise strokes, significantly enhances sketch retrieval performance, and exhibits strong robustness across varying sketch styles.

Zhechun Liang, Tao Huang, Fangfang Wu, Shiwen Xue, Zhenyu Wang, Weisheng Dong, Xin Li, Guangming Shi

Remote sensing composed image retrieval (RSCIR) is a new vision-language task that takes a composed query of an image and text, aiming to search for a target remote sensing image satisfying two conditions from intricate remote sensing imagery. However, the existing attribute-based benchmark Patterncom in RSCIR has significant flaws, including the lack of query text sentences and paired triplets, thus making it unable to evaluate the latest methods. To address this, we propose the Zero-Shot Query Text Generator (ZS-QTG) that can generate full query text sentences based on attributes, and then, by capitalizing on ZS-QTG, we develop the PatternCIR benchmark. PatternCIR rectifies Patterncom’s deficiencies and enables the evaluation of existing methods. Additionally, we explore zero-shot composed image retrieval methods that do not rely on massive pre-collected triplets for training. Existing methods use only the text during retrieval, performing poorly in RSCIR. To improve this, we propose Text-image Sequential Training of Composed Image Retrieval (TisCIR). TisCIR undergoes sequential training of multiple self-masking projection and fine-grained image attention modules, which endows it with the capacity to filter out conflicting information between the image and text, enhancing the retrieval by utilizing both modalities in harmony. TisCIR outperforms existing methods by 12.40% to 62.03% on PatternCIR, achieving state-of-the-art performance in RSCIR. The data and code are available here.

Yongqian Li, Yong Luo, Xiantao Cai, Zheng He, Zhennan Meng, Nidong Wang, Yunlin Chen, Zhifei Li

We proposes a novel visual tokenizer by combining high-level semantic tokens and low-level pixel tokens to represent images, aiming to address the challenges of image-to-sequence conversion for Large Language Models (LLMs). Existing visual tokenizers, such as VQ-VAE and diffusion-based models, either struggle with token explosion as image resolution increases or fail to capture detailed structural information. Our method introduces a dual-token system: high-level semantic tokens capture the main content of the image, while low-level pixel tokens preserve structural details. By integrating these tokens in a hybrid architecture, we leverage a VQ-VAE branch to generate low-resolution guidance and a diffusion process to reconstruct high-resolution images with both semantic coherence and structural accuracy. This approach significantly reduces the number of required tokens and enhances image reconstruction quality, offering an efficient solution for tasks like image generation and understanding based on LLMs.

Yiming Li, Xiaoshan Yang, Bing-Kun Bao, Changsheng Xu

Due to the dynamic nature in videos, it is evident that perceiving and reasoning about temporal information are the key focus of Video Question Answering (VideoQA). In recent years, several methods have explored relationship-level temporal modeling with graph-structured video representation. Unfortunately, these methods heavily rely on the question text, thus making it challenging to perceive and reason about video content that is not explicitly mentioned in the question. To address the above challenge, we propose Graph Prompts-based VideoQA (GP-VQA), which adopts a video-based graph structure for enhanced video understanding. The proposed GP-VQA contains two stages, i.e., pre-training and prompt tuning. In pre-training, we define the pretext task that requires GP-VQA to reason about the randomly masked nodes or edges in the video graph, thus prompting GP-VQA to learn the reasoning ability with video-guided information. In prompt-tuning, we organize the textual question into question graph and implement message passing from video graph to question graph, therefore inheriting the video-based reasoning ability from video graph completion to VideoQA. Extensive experiments on various datasets have demonstrated the promising performance of GP-VQA.

Wei Li, Dezhao Luo, Dongbao Yang, Zhenhang Li, Weiping Wang, Yu Zhou

Video action understanding tasks in real-world scenarios often suffer from data limitations. In this paper, we address the data-limited action understanding problem by bridging data scarcity. We propose a novel method that leverages a text-to-video diffusion transformer to generate annotated data for model training. This paradigm enables the generation of realistic annotated data on an infinite scale without human intervention. We proposed the Information Enhancement Strategy and the Uncertainty-Based Soft Target tailored to generate sample training. Through quantitative and qualitative analyzes, we discovered that real samples generally contain a richer level of information compared to generated samples. Based on this observation, the information enhancement strategy was designed to enhance the informational content of the generated samples from two perspectives: the environment and the character. Furthermore, we observed that a portion of low-quality generated samples might negatively affect model training. To address this, we devised an uncertainty-based label-smoothing strategy to increase the smoothing of these low-quality samples, thereby reducing their impact. We demonstrate the effectiveness of the proposed method on four datasets and five tasks, and achieve state-of-the-art performance for zero-shot action recognition.

Ping Li, Jianan Ni, Bo Pang

Action recognition models using deep learning are vulnerable to adversarial examples, which are transferable across other models trained on the same data modality. Existing transferable attack methods face two major challenges: 1) they heavily rely on the assumption that the decision boundaries of the surrogate (a.k.a., source) model and the target model are similar, which limits the adversarial transferability; and 2) their decision boundary difference makes the attack direction uncertain, which may result in the gradient oscillation, weakening the adversarial attack. This motivates us to propose a Background Mixup-induced Temporal Consistency (BMTC) attack method for action recognition. From the input transformation perspective, we design a model-agnostic background adversarial mixup module to reduce the surrogate-target model dependency. In particular, we randomly sample one video from each category and make its background frame, while selecting the background frame with the top attack ability for mixup with the clean frame by reinforcement learning. Moreover, to ensure an explicit attack direction, we leverage the background category as guidance for updating the gradient of adversarial example, and design a temporal gradient consistency loss, which strengthens the stability of the attack direction on subsequent frames. Empirical studies on two video datasets, i.e., UCF101 and Kinetics-400, and one image dataset, i.e., ImageNet, demonstrate that our method significantly boosts the transferability of adversarial examples across several action/image recognition models.

Dongdong Li, Zhinan Gao, Yangliu Kuai, Rui Chen

Sequence learning based tracking frameworks are popular in the tracking community. In practice, its auto-regressive sequence generation manner leads to inferior performance and high latency compared with latest advanced trackers. In this paper, to mitigate this issue, we propose an efficient and effective sequence-to-sequence tracking framework named FastSeqTrack. FastSeqTrack differs from previous sequence learning based trackers in terms of token initialization and sequence generation manner. Four tracking tokens are appended to patch embeddings and generated in the encoder as initial guesses for the bounding box sequence, which improves the tracking accuracy compared with randomly initialized tokens. Tracking tokens are then parallelly fed into the decoder in a one-pass manner and greatly boost the forward inference speed compared with the auto-regressive manner. Inspired by the early-exit mechanism, we inject internal classifiers after each decoder layer to early terminate forward inference when the softmax confidence is sufficiently reliable. In easy tracking frames, early exits avoid network overthinking and unnecessary computation. Extensive experiments on multiple benchmarks demonstrate that FastSeqTrack runs over 100 fps and showcases superior performance against state-of-the-art trackers. Codes and models are available at https://github.com/vision4drones/FastSeqTrack.

Yudong Jiang, Baohan Xu, Siqian Yang, Mingyu Ying, Jing Liu, Chao Xu, Siqi Wang, Yidi Wu, Bingwen Zhu, Yue Zhang 等

Animation has gained significant interest in the recent film and TV industry. Despite the success of advanced video generation models like Sora, Kling, and CogVideoX in generating natural videos, they lack the same effectiveness in handling animation videos. Evaluating animation video generation is also a great challenge due to its unique artist styles, violating the laws of physics and exaggerated motions. In this paper, we present a comprehensive system, AniSora, designed for animation video generation, which includes a data processing pipeline, a controllable generation model, and an evaluation benchmark. Supported by the data processing pipeline with over 10M high-quality data, the generation model incorporates a spatiotemporal mask module to facilitate key animation production functions such as image-to-video generation, frame interpolation, and localized image-guided animation. We also collect an evaluation benchmark of 948 various animation videos, with specifically developed metrics for animation video generation. Our entire project is publicly available on https://github.com/bilibili/Index-anisora/tree/main

Yuzhu Ji, Mingshan Sun, Jianyang Shi, Xiaoke Jiang, Yiqun Zhang, Haijun Zhang

Category-level object pose estimation is a longstanding and fundamental task crucial for augmented reality and robotic manipulation applications. Existing RGB-based approaches struggle with multi-stage settings and heavily rely on off-the-shelf techniques, such as object detectors, depth estimators, non-differentiable NOCS shape alignment, etc. Extra dependencies lead to the accumulation of errors and complicate the whole pipeline, limiting the deployment of these approaches in practical applications. This paper streamlined an end-to-end framework unifying the single-frame and video-based category-level pose estimation. Specifically, instead of explicitly introducing extra dependencies, the DINOv2 encoder and depth decoder, as robust semantic and geometric prior extractors, are leveraged to produce intra-frame hierarchical semantic and geometric features. A spatial-temporal sparse query network is developed to model the implicit correspondence and inter-frame correlations between a set of implicit 3D query anchors and intra-frame features. Finally, a pose prediction head is employed using the bipartite matching algorithm. Experimental results demonstrate that our model achieves state-of-the-art performance compared with RGB-based categorical pose estimation methods on the REAL275 and CAMERA25 datasets. Our code is available at https://andrewchiyz.github.io/vision.3dv.seqpose/.

Jinglin Huang, Maoqiang Wu, Peichun Li, Wen Wu, Rong Yu

Cross-view geo-localization is a crucial task with diverse applications, yet it remains challenging due to the significant variations in viewpoints and visual appearances between images from different perspectives. While recent advancements have been made, existing methods often suffer from high model complexity, excessive resource consumption, and the impact of sample learning difficulty on optimization. To overcome these limitations, we optimize the Vision Mamba (Vim) model, built on a State Space Model (SSM) architecture, by replacing the traditional classification head with Channel Group Pooling (CGP) for efficient feature integration. This optimization reduces model parameters by 1.5% and computational complexity by 0.4%. Additionally, we propose a novel Dynamic Weighted Batch-tuple Loss (DWBL) to dynamically adjust the weighting of negative samples, improving model performance. By combining CGP and DWBL, we develop an efficient end-to-end network, VimGeo, which achieves state-of-the-art performance with enhanced computational efficiency. Specifically, VimGeo achieves a Recall@1 of 81.67% on the CVACT_test dataset, outperforming prior approaches. Extensive experiments on CVUSA, CVACT, and VIGOR datasets validate VimGeo's effectiveness and competitiveness in cross-view geo-localization tasks, achieving the leading results among sequence modeling-based methods. The implementation is available at: https://github.com/VimGeoTeam/VimGeo.

Feizhen Huang, Yu Wu, Yutian Lin, Bo Du

Video-to-Audio (V2A) Generation achieves significant progress and plays a crucial role in film and video post-production. However, current methods overlook the cinematic language, a critical component of artistic expression in filmmaking. As a result, their performance deteriorates in scenarios where Foley targets are only partially visible. To address this challenge, we propose a simple self-distillation approach to extend V2A models to cinematic language scenarios. By simulating the cinematic language variations, the student model learns to align the video features of training pairs with the same audio-visual correspondences, enabling it to effectively capture the associations between sounds and partial visual information. Our method not only achieves impressive improvements under partial visibility across all evaluation metrics, but also enhances performance on the large-scale V2A dataset, VGGSound.

Mingce Guo, Jingxuan He, Yufei Yin, Zhangye Wang, Shengeng Tang, Lechao Cheng

Text-driven video editing powered by generative diffusion models holds significant promise for applications spanning film production, advertising, and beyond. However, the limited expressiveness of pre-trained word embeddings often restricts nuanced edits, especially when targeting novel concepts with specific attributes. In this work, we present a novel Concept-Augmented Textual Inversion (CATI) framework that flexibly integrates new object information from user-provided concept videos. By fine-tuning only the V (Value) projection in attention via Low-Rank Adaptation (LoRA), our approach preserves the original attention distribution of the diffusion model while efficiently incorporating external concept knowledge. To further stabilize editing results and mitigate the issue of attention dispersion when prompt keywords are modified, we introduce a Dual Prior Supervision (DPS) mechanism. DPS supervises cross-attention between the source and target prompts, preventing undesired changes to non-target areas and improving the fidelity of novel concepts. Extensive evaluations demonstrate that our plug-and-play solution not only maintains spatial and temporal consistency but also outperforms state-of-the-art methods in generating lifelike and stable edited videos. The source code is publicly available at https://guomc9.github.io/STIVE-PAGE/.

Wenxuan Ge, Peng Huang, Rui Yan, Hongyu Qu, Guosen Xie, Xiangbo Shu

Adapting pre-trained vision-language models to downstream tasks has emerged as a novel paradigm for zero-shot learning. Existing test-time adaptation (TTA) methods such as TPT attempt to fine-tune visual or textual representations to accommodate downstream tasks but still require expensive optimization costs. To this end, Training-free Dynamic Adapter (TDA) maintains a cache containing visual features for each category in a parameter-free manner and measures sample confidence based on prediction entropy of test samples. Inspired by TDA, this work aims to develop the first training-free adapter for zero-shot video classification. Capturing the intrinsic temporal relationships within video data to construct and maintain the video cache is key to extending TDA to the video domain. In this work, we propose a reliable and diverse Hierarchical Adapter for zero-shot video classification, which consists of Frame-level Cache Refiner and Video-level Cache Updater. Before each video sample enters the corresponding cache, it needs to be refined at frame level based on prediction entropy and temporal probability difference. Due to the limited capacity of the cache, we update the cache during inference based on the principle of diversity. Experiments on four popular video classification benchmarks demonstrate the effectiveness of Hierarchical Adapter. The code is available at https://github.com/Gwxer/Hierarchical-Adapter.