论文检索

Zicheng Zhao, Kangyu Wang, Shijie Li, Rui Qian, Weiyao Lin, Huabin Liu

Despite advancements in Video Large Language Models (Vid-LLMs) improving multimodal understanding, challenges persist in streaming video reasoning due to its reliance on contextual information. Existing paradigms feed all available historical contextual information into Vid-LLMs, resulting in a significant computational burden for visual data processing. Furthermore, the inclusion of irrelevant context distracts models from key details. This paper introduces a challenging task called Context-guided Streaming Video Reasoning (CogStream), which simulates real-world streaming video scenarios, requiring models to identify the most relevant historical contextual information to deduce answers for questions about the current stream. To support CogStream, we present a densely annotated dataset featuring extensive and hierarchical question-answer pairs, generated by a semi-automatic pipeline. Additionally, we present CogReasoner as a baseline model. It effectively tackles this task by leveraging visual stream compression and historical dialogue retrieval. Extensive experiments prove the effectiveness of this method.

Zhiqian Zhao, Liang Li, Lei Shen, Xichun Sheng, Yaoqi Sun, Fang Kang, Chenggang Yan

Existing text-video retrieval methods mainly focus on singlemodal video content (i.e., visual entities), often overlooking heterogeneous scene text that is ubiquitous in human environments. Although scene text in videos provides finegrained semantics for cross-modal retrieval, effectively utilizing it presents two key challenges: (1) Temporally dense scene text disrupts sync with sparse video frames, obstructing video understanding;(2) Redundant scene text and irrelevant video frames hinder the learning of discriminative temporal clues for retrieval. To address them, we propose a temporal scene-text calibrating and distilling (TCD) network for textvideo retrieval. Specifically, we first design a window-OCR captioner that aggregates dense scene text into OCR captions to facilitate feature interaction. Next, we devise a heterogeneous semantics calibration module that leverages scene text as a self-supervised signal to temporally align window-level OCR captions and frame-level video features. Further, we introduce a context-guided temporal clue distillation module to learn the complementary and relevant details between scene text and video modalities, thereby obtaining discriminative temporal clues for retrieval. Extensive experiments show that our TCD achieves state-of-the-art performance on three scene-text related benchmarks.

Yunpeng Zhao, Cheng Chen, Qing You Pang, Yibing Fu, Quanzheng Li, Carol Tang, Beng Ti Ang, Yueming Jin

Addressing missing modalities is a critical challenge in multimodal brain tumor segmentation. Most existing approaches merely handle modality-incomplete inputs during inference, assuming a full set of modalities for all training samples. However, this unrealistic assumption limits the usage of abundant modality-incomplete data commonly observed in clinical practice. In this paper, we explore a more practical task of tackling missing modalities during both training and inference. We propose a universal model featuring robust modality reconstruction and prompt-guided modality adaptation. Our mask-reconstruction pre-training enables robust modality-invariant representation learning, during which we design a novel distribution approximation method that supervises the reconstruction of absent modalities without requiring full-modal training data. Afterwards, when adapting our model to the segmentation task, we introduce the complete-then-distill (CTD) paradigm, which first estimates missing modalities in training samples from the available ones, and then distills the knowledge from the reconstructed full-modal representations to enhance learning from modality-incomplete data. Moreover, we propose prompt-guided modality adaptation to personalize a subset of model parameters during CTD, enabling the model to adapt to each distinct modality input scenario by using prompts with rich visual-textual information. Extensive experiments on two brain tumor segmentation benchmarks show our method consistently surpasses previous state-of-the-art approaches under dual-stage missing modality settings across various missing ratios.

Yueming Zhao, Hongyu Yang, Di Huang

Generating high-quality, controllable, and structurally consistent 3D scenes in complex multi-object environments remains a fundamental challenge. We present SceneGenesis, a unified framework that synthesizes 3D scenes by combining semantic structural priors with mesh-guided video–geometry fusion. SceneGenesis first employs large language models to convert textual descriptions into category-aware object specifications, which are transformed into structured meshes using procedural approximations and pretrained asset generators, enabling precise layout control and scalable scene construction. To obtain rich and style-controllable appearances, SceneGenesis generates multi-view video representations conditioned on the initialized structure. A mesh-guided video–geometry fusion module then consolidates video evidence with mesh priors through mesh-conditioned fragment initialization, progressive geometric refinement, and structure-aware optimization, substantially improving global geometric fidelity and visual realism. Experiments demonstrate that SceneGenesis supports flexible style variation and object-level editing while achieving strong controllability, scalability, and structural quality.

Yuanzhi Zhao, Fan Yang, Yudong Zhao, Xiaoyu Li

In cross-modal retrieval tasks, unsupervised hash code learning still faces key challenges, including the difficulty of modeling shared semantic structures across modalities and the inability to adaptively balance multiple supervision objectives during optimization. To address these issues, we propose a novel Unsupervised Dynamic Weighted Cluster-Cooperative Hashing (UDCH) framework, which jointly models feature-level alignment and cluster-level semantic structure to guide consistency learning across modalities under label-free conditions. Specifically, we design an instance-level contrastive loss in the feature branch to align the embedding spaces of images and texts, while employing K-Means clustering to generate pseudo-labels and construct a cluster-center contrast mechanism that captures semantic grouping information. Furthermore, we integrate cross-modal feature similarity to construct a high-order structure matrix, enabling fine-grained structural supervision. To enhance the synergy of multi-objective optimization, we introduce a dynamic weighting strategy that adaptively adjusts the contributions of the feature and cluster branches based on the degree of modal alignment and semantic compactness. Extensive experiments on multiple cross-modal retrieval benchmarks demonstrate that UDCH achieves superior semantic alignment and retrieval performance under unsupervised settings, validating the effectiveness of multi-level semantic modeling and adaptive collaboration mechanisms in unsupervised hashing tasks.

Youjun Zhao, Jiaying Lin, Shuquan Ye, Qianshi Pang, Rynson W. H. Lau

Open-vocabulary 3D scene understanding (OV-3D) aims to localize and classify novel objects beyond the closed set of object classes. However, existing approaches and benchmarks primarily focus on the open vocabulary problem within the context of object classes, which is insufficient in providing a holistic evaluation to what extent a model understands the 3D scene. In this paper, we introduce a more challenging task called Generalized Open-Vocabulary 3D Scene Understanding (GOV-3D) to explore the open vocabulary problem beyond object classes. It encompasses an open and diverse set of generalized knowledge, expressed as linguistic queries of fine-grained and object-specific attributes. To this end, we contribute a new benchmark named OpenScan, which consists of 3D object attributes across eight representative linguistic aspects, including affordance, property, and material. We further evaluate state-of-the-art OV-3D methods on our OpenScan benchmark and discover that these methods struggle to comprehend the abstract vocabularies of the GOV-3D task, a challenge that cannot be addressed simply by scaling up object classes during training. We highlight the limitations of existing methodologies and explore promising directions to overcome the identified shortcomings.

Xuan Zhao, Zhongyu Zhang, Yuge Huang, Yuxi Mi, Guodong Mu, Shouhong Ding, Jun Wang, Rizen Guo, Shuigeng Zhou

Existing state-of-the-art image tokenization methods leverage diverse semantic features from pre-trained vision models for additional supervision, to expand the distribution of latent representations and thereby improve the quality of image reconstruction and generation. These methods employ a locally supervised approach for semantic supervision, which limits the uniformity of semantic distribution. However, VA-VAE proves that a more uniform feature distribution yields better generation performance. In this work, we introduce a Global Perspective Tokenizer (GloTok), which utilizes global relational information to model a more uniform semantic distribution of tokenized features. Specifically, a codebook-wise histogram relation learning method is proposed to transfer the semantics, which are modeled by pre-trained models on the entire dataset, to the semantic codebook. Then, we design a residual learning module which recovers the fine-grained details to minimize the reconstruction error caused by quantization. Through the above design, GloTok delivers more uniformly distributed semantic latent representations, which facilitates the training of autoregressive (AR) models for generating high-quality images without requiring direct access to pre-trained models during the training process. Experiments on the standard ImageNet-1k benchmark clearly show that our proposed method achieves state-of-the-art reconstruction performance and generation quality.

Xiaoming Zhao, Alex Schwing

Classifier-free guidance has become a staple for conditional generation with denoising diffusion models. However, a comprehensive understanding of classifier-free guidance is still missing. In this work, we carry out an empirical study to provide a fresh perspective on classifier-free guidance. Concretely, instead of solely focusing on classifier-free guidance, we trace back to the root, i.e., classifier guidance, pinpoint the key assumption for the derivation, and conduct a systematic study to understand the role of the classifier. On 1D data, we find that both classifier guidance and classifier-free guidance achieve conditional generation by pushing the denoising diffusion trajectories away from decision boundaries, i.e., areas where conditional information is usually entangled and is hard to learn. To validate this classifier-centric perspective on high-dimensional data, we assess whether a flow-matching postprocessing step that is designed to narrow the gap between a pre-trained diffusion model’s learned distribution and the real data distribution, especially near decision boundaries, can improve the performance. Experiments on various datasets verify our classifier-centric understanding.

Xiaojie Zhao, Jinqiao Shi, Yi Li, Junmin Huang, Chongru Fan

Federated learning (FL) allows for collaborative model training while preserving data privacy, but its distributed nature makes it vulnerable to poisoning attacks. Existing defense methods typically rely on using gradients from multiple clients to define a trusted region, selecting only the trustworthy update (good gradients) within this region for aggregation. Mainstream defense boundaries are categorized as hard boundaries, soft boundaries, and semi-soft boundaries. However, we argue that even good gradients within these boundaries can still be exploited by attackers to poison the model. To tackle this challenge, we introduce a boundary-adaptive attack method that leverages the directional properties of optimization techniques to derive baseline poisoned gradients. Through iterative perturbation, it generates seemingly innocent gradients that subtly deviate from the global model. Our extensive study on benchmark datasets and mainstream defensive mechanisms confirms that the proposed attack raises a significantly threat to the integrity and security of FL practices, regardless of the flourishing of robust FL methods.

Weiyu Zhao, Chenyang Wang, Liangxiao Hu, Zonglin Li, Wei Yu, Shengping Zhang

We propose DialoGen, a novel framework for generating realistic gestures for both interlocutors in dialog scenarios, conditioned on conversational audios. Unlike most existing methods that focus solely on a single speaker, DialoGen simultaneously generates synchronized gestures for both participants while also embedding identity-decoupled style into generated gestures that enhance realism and expressiveness. To ensure precise synchronization between interlocutors, DialoGen adopts an interactive dual-diffusion model with mutual interaction estimation, which integrates interaction correlation into the diffusion process. More importantly, by leveraging supervised contrastive learning, we develop the identity-decoupled style guidance to adaptively decompose the identity-specific style of interlocutors into latent space, enabling multi-style dialog gesture generation. Extensive experimental results demonstrate that our model significantly outperforms existing methods in generating realistic, speech-aligned, identity-specific gestures, offering a high-quality solution for various dialog scenarios.

Shuting Zhao, Zeyu Xiao, Xinrong Chen

Full-body motion tracking plays an essential role in AR/VR applications, bridging physical and virtual interactions. However, it is challenging to reconstruct realistic and diverse full-body poses based on sparse signals obtained by head-mounted displays, which are the main devices in AR/VR scenarios. Existing methods for pose reconstruction often incur high computational costs or rely on separately modeling spatial and temporal dependencies, making it difficult to balance accuracy, temporal coherence, and efficiency. To address this problem, we propose KineST, a novel kinematics-guided state space model, which effectively extracts spatiotemporal dependencies while integrating local and global pose perception. The innovation comes from two core ideas. Firstly, in order to better capture intricate joint relationships, the scanning strategy within the State Space Duality framework is reformulated into kinematics-guided bidirectional scanning, which embeds kinematic priors. Secondly, a mixed spatiotemporal representation learning approach is employed to tightly couple spatial and temporal contexts, balancing accuracy and smoothness. Additionally, a geometric angular velocity loss is introduced to impose physically meaningful constraints on rotational variations for further improving motion stability. Extensive experiments demonstrate that KineST has superior performance in both accuracy and temporal consistency within a lightweight framework.

Shijie Zhao, Zhenyu Liang, Xing Yang, Haoqi Gao, Anjie Peng, Hui Zeng

Unrestricted adversarial attacks aim to fool DNNs by generating effective yet photorealistic examples. However, previous methods usually rely on global perturbations to enhance attack performance, which inevitably introduces visual distortions. To reduce visual distortions in the background, we propose a diffusion-based framework that focuses on local perturbations to generate object-level unrestricted adversarial examples (ObjectAdv). Since the cross-attention maps of Stable Diffusion contain the object information, we directly leverage the attention maps to localize the semantic region of object where for attacking. Second, a prompt-switching strategy is proposed for both imperceptibility and attack capacity. Specifically, to preserve layout and object shape of clean image, a prompt of true category is used at early denoising steps. At the later steps, we propose a well-designed prompt to guide the diffusion model to generate transferable adversarial examples. This local attack may cause inconsistency between the perturbed object and the background in adversarial examples. An FFT-based edge smoother is utilized to ensure seamless blending of the edges. ObjectAdv achieves an average ASR of 99.2% in white-box test on the ImageNet-compatible dataset, and outperforms existing methods on defense performance (+5%) and image quality metrics, e.g., SSIM of 0.9140 (+0.1048) and FID of 25.63 (-19.27).

Shenao Zhao, Pengpeng Liang, Zhoufan Yang

Unsupervised domain adaptation for LiDAR-based 3D object detection (3D UDA) based on the teacher-student architecture with pseudo labels has achieved notable improvements in recent years. Although it is quite popular to collect point clouds and images simultaneously, little attention has been paid to the usefulness of image data in 3D UDA when training the models. In this paper, we propose an approach named MMAssist that improves the performance of 3D UDA with multi-modal assistance. A method is designed to align 3D features between the source domain and the target domain by using image and text features as bridges. More specifically, we project the ground truth labels or pseudo labels to the images to get a set of 2D bounding boxes. For each 2D box, we extract its image feature from a pre-trained vision backbone. A large vision-language model (LVLM) is adopted to extract the box's text description, and a pre-trained text encoder is used to obtain its text feature. During the training of the model in the source domain and the student model in the target domain, we align the 3D features of the predicted boxes with their corresponding image and text features, and the 3D features and the aligned features are fused with learned weights for the final prediction. The features between the student branch and the teacher branch in the target domain are aligned as well. To enhance the pseudo labels, we use an off-the-shelf 2D object detector to generate 2D bounding boxes from images and estimate their corresponding 3D boxes with the aid of point cloud, and these 3D boxes are combined with the pseudo labels generated by the teacher model. Experimental results show that our approach achieves promising performance compared with state-of-the-art methods in three domain adaptation tasks on three popular 3D object detection datasets.

Rui Zhao, Shuoyao Wang, Xinhu Zheng, Shijian Gao

Multi-view 3D object detection has garnered increasing attention, particularly due to its success in autonomous driving systems. Although multi-view systems possess rich semantic information, their spatial-geometric reasoning capabilities remain limited. Recent studies employ simulated point cloud generation mechanisms to facilitate LiDAR-camera multi-modal knowledge distillation, achieving formal structural consistency. Despite advancements, these methods still face two main issues: i) alignment challenges caused by discrepancies between LiDAR and camera data, and ii) prediction errors from simulated point clouds that compromise the semantic information extracted from images during fusion. To address these problems, we propose adaptive-smooth distillation to optimize alignment granularity based on feature discrepancies for improved LiDAR-camera knowledge distillation. Specifically, this work considers both LIDAR-to-camera cross-modal distillation and LiDAR-camera fusion to simulated point cloud-camera fusion multi-modal distillation. Then, we introduce a heterogeneous fusion module to strategically bias the fusion process toward the extracted camera features, thereby enhancing the robustness of the fusion feature. Additionally, soft-weighted response distillation is proposed to facilitate the student model to selectively mimic the high-quality output of the teacher model. Extensive experiments have demonstrated the superiority of our method, achieving statistically significant improvements of 4.9% in mean Average Precision (mAP) and 4.5% in NuScenes Detection Score (NDS) over the benchmark.

Rongzhen Zhao, Jian Li, Juho Kannala, Joni Pajarinen

Unsupervised video Object-Centric Learning (OCL) is promising as it enables object-level scene representation and understanding as we humans do. Mainstream video OCL methods adopt a recurrent architecture: An aggregator aggregates current video frame into object features, termed slots, under some queries; A transitioner transits current slots to queries for the next frame. This is an effective architecture but all existing implementations both (i1) neglect to incorporate next frame features, the most informative source for query prediction, and (i2) fail to learn transition dynamics, the knowledge essential for query prediction. To address these issues, we propose Random Slot-Feature pair for learning Query prediction (RandSF.Q): (t1) We design a new transitioner to incorporate both slots and features, which provides more information for query prediction; (t2) We train the transitioner to predict queries from slot-feature pairs randomly sampled from available recurrences, which drives it to learn transition dynamics. Experiments on scene representation demonstrate that our method surpass existing video OCL methods significantly, e.g., up to 10 points on object discovery, setting new state-of-the-art. Such superiority also benefits downstream tasks like scene understanding.

Libo Zhao, Xiaoli Zhang, Zeyu Wang

Infrared and Visible Image Fusion (IVIF) produces enhanced images by fusing complementary visual information. However, most existing methods generate fixed outputs and cannot flexibly adapt to user-specific requirements. Recent text-guided approaches offer partial control but are limited to global or semantic levels, lacking instance-level control. This limitation arises from two challenges: first, the lack of datasets that directly link textual instructions with corresponding spatial annotations, and second, the use of coarse cross-modal alignment methods that struggle to precisely match textual instructions with visual features. To overcome these challenges, we propose ControlFuse, a controllable IVIF framework enabling multi-granularity fusion across global, semantic, and instance levels, guided by user instructions. First, we construct an automated multi-granularity dataset that provides explicit textual-mask correspondences at these three levels. Second, inspired by manifold geometry, we design a Multimodal Feature Interaction Module (MFIM) comprising Feature Manifold Converter (FMC) and Curvature-Guided Interaction (CGI). FMC projects textual and visual features into a unified manifold space, while CGI leverages manifold curvature as a geometric cue to refine cross-modal alignment. Extensive experiments validate ControlFuse, outperforming state-of-the-art methods in robustness and flexibility.

Kangqiao Zhao, Shuo Huai, Xurui Song, Jun Luo

Though deep neural models adopted to realize the perception of autonomous driving have proven vulnerable to adversarial examples, known attacks often leverage 2D patches and target mostly monocular perception. Therefore, the effectiveness of Physical Adversarial Examples (PAEs) on stereo-based binocular depth estimation remains largely unexplored. To this end, we propose the first texture-enabled physical adversarial attack against stereo matching models in the context of autonomous driving. Our method employs a 3D PAE with global camouflage texture rather than a local 2D patch-based one, ensuring both visual consistency and attack effectiveness across different viewpoints of stereo cameras. To cope with the disparity effect of these cameras, we also propose a new 3D stereo matching rendering module that allows the PAE to be aligned with real-world positions and headings in binocular vision. We further propose a novel merging attack that seamlessly blends the target into the environment through fine-grained PAE optimization. It has significantly enhanced stealth and lethality upon existing hiding attacks that fail to get seamlessly merged into the background. Extensive evaluations show that our PAEs can successfully fool the stereo models into producing erroneous depth information.

Kai Zhao, Liting Ruan, Haoran Jiang, Xiaoqiang Zhu, Xianchao Zhang, Dan Zeng

Images are typically sampled on a uniform grid,despite their non-uniform information distribution—some regions are rich in content while others are not. The mismatch leads to inefficient computation allocation in deep learning models. To address this, recent studies have proposed predictive downsampling methodsthat adaptively downsample images based on predicted per-pixel importance, allocating more pixels to informative areas. However,these methods require high-resolution processing to accurately estimate importance, which undermines their efficiency:the prediction itself must process the full-resolution image,consuming most of the computational budget. This high-resolution importance prediction is necessary because each input may differ significantly in structure and content. In this paper, we take a different approach and introduce a learn-to-downsample paradigmtailored for aligned vision recognition tasks, such as face recognition and palmprint recognition, where input alignment ensures consistent spatial structure across images. This alignment ensures structural consistency across images, allowing a shared, input-agnostic downsampling template applicable to all inputs. Furthermore, instead of relying on implicit importance maps, we introduce a flow-based representation that explicitly models the spatial warping from the original image to the downsampled version. The flow representation is not only more efficient but also more controllable: we regularize the flow using its Jacobian determinant to precisely control the sampling density and coverage,enabling interpretable and tunable sampling patterns. Extensive experiments on two aligned recognition tasks, face and palmprint recognition, demonstrate that our method substantially reduces computational cost with minimal accuracy degradation, achieving a significantly better performance-efficiency trade-off than existing predictive downsampling methods.

Jinghan Zhao, Yifei Huang, Feng Lu

Learning procedural-aware video representations is a key step towards building agents that can reason about and execute complex tasks. Existing methods typically address this problem by aligning visual content with textual descriptions at the task and step levels to inject procedural semantics into video representations. However, due to their high level of abstraction, "task" and "step" descriptions fail to form a robust alignment with the concrete, observable details in visual data. To address this, we introduce "states", i.e., textual snapshots of object configurations, as a visually-grounded semantic layer that anchors abstract procedures to what a model can actually see. We formalize this insight in a novel Task-Step-State (TSS) framework, where tasks are achieved via steps that drive transitions between observable states. To enforce this structure, we propose a progressive pre-training strategy that unfolds the TSS hierarchy, forcing the model to first ground representations in states before associating them with steps and, ultimately, high-level tasks. Extensive experiments on the COIN and CrossTask datasets show that our method outperforms baseline models on multiple downstream tasks, including task recognition, step recognition, and next step prediction. Ablation studies show that introducing state supervision is a key driver of performance gains across all tasks. Additionally, our progressive pretraining strategy proves more effective than standard joint training, as it better enforces the intended hierarchical structure.

Jiaxing Zhao, Boyuan Sun, Xiang Chen, Xihan Wei

Facial expression captioning has found widespread application across various domains. Recently, the emergence of video Multimodal Large Language Models (MLLMs) has shown promise in general video understanding tasks. However, describing facial expressions within videos poses two major challenges for these models: (1) the lack of adequate datasets and benchmarks, and (2) the limited visual token capacity of video MLLMs. To address these issues, this paper introduces a new instruction-following dataset tailored for dynamic facial expression caption. The dataset comprises 5,033 high-quality video clips annotated manually, containing over 700,000 tokens. Its purpose is to improve the capability of video MLLMs to discern subtle facial nuances. Furthermore, we propose FaceTrack-MM, which leverages a limited number of tokens to encode the main character’s face. This model demonstrates superior performance in tracking faces and focusing on the facial expressions of the main characters, even in intricate multiperson scenarios. Additionally, we introduce a novel evaluation metric combining event extraction, relation classification, and the longest common subsequence (LCS) algorithm to assess the content consistency and temporal sequence consistency of generated text. Moreover, we present FECBench, a benchmark designed to assess the performance of existing video MLLMs in this specific task.