Multimodal Large Language Models (MLLMs) possess extensive knowledge and strong reasoning capabilities, achieving remarkable performance in knowledge-based visual question answering, significantly surpassing traditional small-scale Vision-Language Models (VLMs). However, the distinct training paradigms of MLLMs and small-scale VLMs result in misaligned feature representation spaces and divergent answer prediction distributions. To bridge this gap, we propose a novel end-to-end large-small model synergy framework, where small VLMs and MLLMs collaborate via synergistic optimization of shared objectives while maintaining their co-evolving complementary specializations. Specifically, multimodal fine-grained heuristics are extracted from well-tuned small VLMs and subsequently projected into the textual space of MLLMs through dedicated visual and textual collaboration modules. This enables cross-modal guidance for both visual and textual inputs. Finally, a dual-objective synergy loss promotes alignment toward shared goals, while a visual discrepancy loss preserves specialization diversity. Extensive experiments demonstrate that our framework achieves state-of-the-art performance on both the OK-VQA and A-OKVQA benchmarks.
论文检索
输入标题、作者或关键词，从 1,620 篇学术成果中精准定位
Online 3D occupancy prediction provides a comprehensive spatial understanding of embodied environments. While the innovative EmbodiedOcc framework utilizes 3D semantic Gaussians for progressive indoor occupancy prediction, it overlooks the geometric characteristics of indoor environments, which are primarily characterized by planar structures. This paper introduces EmbodiedOcc++, enhancing the original framework with two key innovations: a Geometry-guided Refinement Module (GRM) that constrains Gaussian updates through plane regularization, along with a Semantic-aware Uncertainty Sampler (SUS) that enables more effective updates in overlapping regions between consecutive frames. GRM regularizes the position update to align with surface normals. It determines the adaptive regularization weight using curvature-based and depth-based constraints, allowing semantic Gaussians to align accurately with planar surfaces while adapting in complex regions. To effectively improve geometric consistency from different views, SUS adaptively selects proper Gaussians to update. Comprehensive experiments on the EmbodiedOcc-ScanNet benchmark demonstrate that EmbodiedOcc++ achieves state-of-the-art performance across different settings. Our method demonstrates improved edge accuracy and retains more geometric details while ensuring computational efficiency, which is essential for online embodied perception. The code will be released at: https://github.com/PKUHaoWang/EmbodiedOcc2.
Due to the frequent occurrence of missing views in real-world multi-view data, incomplete multi-view clustering (IMVC) has attracted significant attention. However, most existing IMVC methods overlook the fact that incomplete data in practical applications often exhibits varying missing rates across different views, rendering their mechanisms ineffective under such conditions. Although several works based on conventional learning methods have been proposed to solve unbalanced incomplete multi-view clustering (UIMVC), their performance is limited by their shallow feature representation and over-sophisticated optimization procedure. In this paper, we propose Deep Unbalanced Incomplete Multi-view Clustering via Graph Constrained Imputation and Contrastive Learning (DUIMC) to address UIMVC with deep learning paradigm. Specifically, DUIMC introduces a novel differentiable imputation layer for dynamically handling unbalanced incompleteness and integrates it with multi-view contrastive clustering into a unified deep representation learning framework. Furthermore, bi-level graph constraints are imposed on imputation and representation learning to preserve local consistency at both the feature and instance levels. In addition, we develop adaptive fusion mechanisms to adaptively restrain the impact aroused by information unbalance among views. Extensive experimental results on five benchmark datasets demonstrate DUIMC's superior clustering performance over several traditional state-of-the-art approaches.
Movie Dubbing aims to convert scripts into speeches that align with the given movie clip in both temporal and emotional aspects while preserving the vocal timbre of a given brief reference audio. Existing methods focus primarily on reducing the word error rate while ignoring the importance of lip-sync and acoustic quality. To address these issues, we propose a novel dubbing architecture based on Large Language Model (LLM) and Conditional Flow Matching (CFM), named FlowDubber, which achieves high-quality audio-visual sync and pronunciation by incorporating a large speech language model with dual contrastive alignment while improving acoustic quality via Flow-based Voice Enhancing (FVE). First, we introduce Qwen2.5 as the backbone of large speech language model to learn the in-context sequence from movie scripts and reference audio. Second, the proposed semantic-aware learning focuses on capturing LLM semantic knowledge at the phoneme level, which facilitates mutual alignment with lip movement from silent video via Dual Contrastive Alignment (DCA). Third, the FVE introduces an LLM-based acoustics flow matching guidance to strengthen clarity by decoupling Classifier-Free Guidance (CFG) enhancement. Extensive experiments demonstrate that our method outperforms several state-of-the-art methods on two primary benchmarks. The demos are available at https://galaxycong.github.io/LLM-Flow-Dubber/.
Existing LiDAR-based 3D object detectors typically rely on manually annotated labels for training to achieve good performance. However, obtaining high-quality 3D labels is time-consuming and labor-intensive. To address this issue, recent works explore unsupervised 3D object detection by introducing RGB images as an auxiliary modal to assist pseudo-box generation. However, these methods simply integrate pseudo-boxes generated by LiDAR point clouds and RGB images. Yet, such a label-level fusion strategy brings limited improvements to the quality of pseudo-boxes, as it overlooks the complementary nature in terms of LiDAR and RGB image data. To overcome the above limitations, we propose a novel data-level fusion framework that integrates RGB images and LiDAR data at an early stage. Specifically, we utilize vision foundation models for instance segmentation and depth estimation on images and introduce a bi-directional fusion method, where real points acquire category labels from the 2D space, while 2D pixels are projected onto 3D to enhance real point density. To mitigate noise from depth and segmentation estimations, we propose a local and global filtering method, which applies local radius filtering to suppress depth estimation errors and global statistical filtering to remove segmentation-induced outliers. Furthermore, we propose a data-level fusion based dynamic self-evolution strategy, which iteratively refines pseudo-boxes under a dense representation, significantly improving localization accuracy. Extensive experiments on the nuScenes dataset demonstrate that the detector trained by our method significantly outperforms that trained by previous state-of-the-art methods with 28.4% mAP on the nuScenes validation benchmark.
Given the scarcity of real data and the time-intensive nature of labeling, current multi-agent perception models often rely on simulated sensor data for training and validation. However, perception performance deteriorates significantly due to domain gap between simulated and real data. Existing adaptation methods focus on domain-generalized feature extraction while neglecting multi-agent shift uncertainty and relational semantic loss. To address this issue, we propose a Selective Shift Domain Adaptation method in multi-agent collaborative perception, called SSDA. SSDA incorporates two essential components: the frequency-decoupled feature shift adjustment (FSA) and the entropy-driven staged adaptive alignment (SAA). To mitigate the relational semantic loss, the FSA is proposed to simplify the representation of correlation features and remove redundant information from the source domain, thereby mitigating interference for domain adversarial scenarios. To tackle the shift uncertainty, the SAA is designed to achieve adaptive alignment from global to local guided by information entropy, which dynamically adjusts weights for samples according to their level of uncertainty. The results demonstrate that the SSDA is significantly superior to the SOTA, achieving up to 7.35% improvements on AP@0.7.
Video text-based visual question answering (Video TextVQA) aims to answer questions by explicitly reading and reasoning about the text involved in a video. Most works in this field follow a frame-level framework which suffers from redundant text entities and implicit relation modeling, resulting in limitations in both accuracy and efficiency. In this paper, we rethink the Video TextVQA task from an instance-oriented perspective and propose a novel model termed GAT (Gather and Trace). First, to obtain accurate reading result for each video text instance, a context-aggregated instance gathering module is designed to integrate the visual appearance, layout characteristics, and textual contents of the related entities into a unified textual representation. Then, to capture dynamic evolution of text in the video flow, an instance-focused trajectory tracing module is utilized to establish spatio-temporal relationships between instances and infer the final answer. Extensive experiments on several public Video TextVQA datasets validate the effectiveness and generalization of our framework. GAT outperforms existing Video TextVQA methods, video-language pretraining methods, and video large language models in both accuracy and inference speed. Notably, GAT surpasses the previous state-of-the-art Video TextVQA methods by 3.86% in accuracy and achieves ten times of faster inference speed than video large language models. The source code is available at https://github.com/zhangyan-ucas/GAT.
Technical Element Score (TES) and Program Component Score (PCS) evaluations in figure skating demand precise assessment of athletic actions and artistic interpretation, respectively. Existing methods face three major challenges. Firstly, video and audio cues are regarded as common features for both TES and PCS predictions in previous works without considering the prior evaluation criterion of figure skating. Secondly, action elements in competitions are separated in time, TES should be derived from each element's score, but existing methods try to give an overall TES prediction without evaluating each action element. Thirdly, lengthy competition videos make it difficult and inefficient to handle long-range contexts. To address these challenges, we propose a two-stream Mamba pyramid network that aligns with actual judging criteria to predict TES and PCS by separating visual-feature based TES evaluation stream from audio-visual-feature based PCS evaluation stream. In the PCS evaluation stream, we introduce a multi-level fusion mechanism to guarantee that video-based features remain unaffected when assessing TES, and enhance PCS estimation by fusing visual and auditory cues across each contextual level of the pyramid. In the TES evaluation stream, the multi-scale Mamba pyramid and TES head we proposed effectively address the challenges of localizing and evaluating action elements with various temporal scales and give score predictions. With Mamba's superior ability to capture long-range dependencies and its linear computational complexity, our method is ideal for handling lengthy figure skating videos. Comprehensive experimentation demonstrates that our framework attains state-of-the-art performance on the FineFS benchmark. Furthermore, it yields competitive outcomes on two additional datasets without further training. Our source code is available at https://github.com/ycwfs/Figure-Skating-Action-Quality-Assessment.
In multi-view clustering (MVC), anchor technique is generally hailed as an effective means for filtering noise and improving computation efficiency. However, existing methods usually construct anchors via heuristic strategy, random sampling, or orthogonal learning, which overlook the distribution differences between anchors and original data, leading to anchors lacking structural characteristics. To generate the anchors that are with similar distributions to original data, in the paper we carefully devise a LASD algorithm from the perspective of optimal transport (OT). Concretely, we firstly design a Multi-View OT (MVOT) framework through complementary and consensus representation learning. Then, we theoretically demonstrate the convexity of MVOT using the positive semidefiniteness of its Hessian matrix, and accordingly the global optimal solution of each transport plan can be reached. Further, we establish the strong dual condition for MVOT by the relative interior. Based on dual programming, consequently, we successfully obtain the transport plan between anchors and original data for each view within linear computational complexity. Afterwards, the spectral clustering operation is employed on the consensus plan to produce the discrete cluster labels. Abundant experiments underscore that our learned anchors do well reflect the distributions of original data, and the generated clustering results outperform multiple strong MVC competitors, even under large-scale scenarios. The source code is available at https://github.com/junpuzhang/LASD.
Generating accurate descriptions of human actions in videos remains a challenging task for video captioning models. Existing approaches often struggle to capture fine-grained motion details, resulting in vague or semantically inconsistent captions. In this work, we introduce the Motion-Augmented Caption Model (M-ACM), a novel generative framework that enhances caption quality by incorporating motion-aware decoding. At its core, M-ACM leverages motion representations derived from human mesh recovery to explicitly highlight human body dynamics, thereby reducing hallucinations and improving both semantic fidelity and spatial alignment in the generated captions. To support research in this area, we present the Human Motion Insight (HMI) Dataset, comprising 115K video-description pairs focused on human movement, along with HMI-Bench, a dedicated benchmark for evaluating motion-focused video captioning. Experimental results demonstrate that M-ACM significantly outperforms previous methods in accurately describing complex human motions and subtle temporal variations, setting a new standard for motion-centric video captioning.
Cross-video question answering presents significant challenges beyond traditional single-video understanding, particularly in establishing meaningful connections across video streams and managing the complexity of multi-source information retrieval. We introduce VideoForest, a novel framework that addresses these challenges through person-anchored hierarchical reasoning, enabling effective cross-video understanding without requiring end-to-end training. VideoForest integrates three key innovations: 1) a human-anchored feature extraction mechanism that employs ReID and tracking algorithms to establish robust spatiotemporal relationships across multiple video sources; 2) a multi-granularity spanning tree structure that hierarchically organizes visual content around person-level trajectories; and 3) a multi-agent reasoning framework that efficiently traverses this hierarchical structure to answer complex queries. To evaluate our method, we develop CrossVideoQA, a comprehensive benchmark specifically designed for person-centric cross-video analysis. Experimental results demonstrate VideoForest's superior performance in cross-video reasoning tasks, achieving 71.93% accuracy in person recognition, 83.75% in behavior analysis, and 51.67% in summarization and reasoning.
Symbolic music understanding is a fundamental task in multimedia interpretation, which aims to decode musical attributes from symbolic music representations. Existing methods usually handle musical sequences as linguistic data, ignoring intrinsic musical properties. For instance, recurring melodies often appear in musical pieces with subtle variations, which requires methods that are aware of both local musical details and overall repetitive patterns, yet current methods can not fit the request. To address this issue, we introduce FG-Midiformer, a modified transformer framework that incorporates a multi-scale-aware feature learning (MSAFL) module and a local feature enhanced classification (LFEC) module for fine-grained understanding of multi-attributes. Specifically, the MSAFL module is designed to capture multi-scale musical relationships by embedding efficient multi-scale attention for long-term dependency modeling. In order to improve the classification accuracy of musical attributes, we devise the LFEC module, in which an attention mechanism with full 3-D weights is first introduced to efficiently highlight and leverage important local musical features. The LFEC module strengthens local feature representation and improves the sensitivity of FG-Midiformer to subtle differences between musical attributes. Extensive experiments show that FG-Midiformer achieves state-of-the-art performance in multi-attribute understanding tasks such as melody identification, velocity prediction, composer categorization, and emotion classification. The code will be released at https://github.com/Viki66666/FG-Midiformer.
Recent advances in diffusion models have notably enhanced text-to-image (T2I) generation quality, but they also raise the risk of generating unsafe content. Traditional safety methods like text blacklisting or harmful content classification have significant drawbacks: they can be easily circumvented or require extensive datasets and extra training. To overcome these challenges, we introduce PurifyGen, a novel, training-free approach for safe T2I generation that retains the model's original weights. PurifyGen introduces a dual-stage strategy for prompt purification. First, we evaluate the safety of each token in a prompt by computing its complementary semantic distance, which measures the semantic proximity between the prompt tokens and concept embeddings from predefined toxic and clean lists. This enables fine-grained prompt classification without explicit keyword matching or retraining. Tokens closer to toxic concepts are flagged as risky. Second, for risky prompts, we apply a dual-space transformation: we project toxic-aligned embeddings into the null space of the toxic concept matrix, effectively removing harmful semantic components, and simultaneously align them into the range space of clean concepts. This dual alignment purifies risky prompts by both subtracting unsafe semantics and reinforcing safe ones, while retaining the original intent and coherence. We further define a token-wise strategy to selectively replace only risky token embeddings, ensuring minimal disruption to safe content. PurifyGen offers a plug-and-play solution with theoretical grounding and strong generalization to unseen prompts and models. Extensive testing shows that PurifyGen surpasses current methods in reducing unsafe content across five datasets and competes well with training-dependent approaches.
The RAW domain image super-resolution faces two critical challenges: the physical impossibility of capturing native high-quality RAW references with a resolution-limited camera and the limitations of neural networks, including inefficient residual layer utilization and spectral bias in feature learning. This paper proposes a strategy combining physics-based imaging simulation and neural networks to jointly address these challenges. First, we develop a rapid imaging simulation system based on our proposed subgraph decomposition technology. It generates camera-specific degraded and clean RAW image pairs at multiple resolutions. Second, we design a LatentKAN network, featuring an iterative feature fusion network that extracts additional beneficial information through stage-wise supervision and a multi-layer Kolmogorov Arnold network that suppresses spectral bias via learnable activation functions. Ultimately, our strategy demonstrates significant advantages, achieving an average 0.8 dB PSNR improvement across all SR scales compared to state-of-the-art methods, thereby establishing a new paradigm for camera-specific super-resolution tasks.
Shallow Features Matter: Hierarchical Memory with Heterogeneous Interaction for Unsupervised Video Object Segmentation
PDF ↗Unsupervised Video Object Segmentation (UVOS) aims to predict pixel-level masks for the most salient objects in videos without any prior annotations. While memory mechanisms have been proven critical in various video segmentation paradigms, their application in UVOS yield only marginal performance gains despite sophisticated design. Our analysis reveals a simple but fundamental flaw in existing methods: over-reliance on memorizing high-level semantic features. UVOS inherently suffers from the deficiency of lacking fine-grained information due to the absence of pixel-level prior knowledge. Consequently, memory design relying solely on high-level features, which predominantly capture abstract semantic cues, is insufficient to generate precise predictions. To resolve this fundamental issue, we propose a novel hierarchical memory architecture to incorporate both shallow- and high-level features for memory, which leverages the complementary benefits of pixel and semantic information. Furthermore, to balance the simultaneous utilization of the pixel and semantic memory features, we propose a heterogeneous interaction mechanism to perform pixel-semantic mutual interactions, which explicitly considers their inherent feature discrepancies. Through the design of Pixel-guided Local Alignment Module (PLAM) and Semantic-guided Global Integration Module (SGIM), we achieve delicate integration of the fine-grained details in shallow-level memory and the semantic representations in high-level memory. Our Hierarchical Memory with Heterogeneous Interaction Network (HMHI-Net) consistently achieves state-of-the-art performance across all UVOS and video saliency detection benchmarks. Moreover, HMHI-Net consistently exhibits high performance across different backbones, further demonstrating its superiority and robustness. Project page: https://github.com/ZhengxyFlow/HMHI-Net.
Given that action evolution follows temporal progression, recent studies for Online Action Detection (OAD) and Online Action Anticipation (OAA) generally adopt forward temporal modeling to capture dependencies in observable video sequences. However, the strictly sequential nature of forward temporal modeling prevents subsequent frames from being used to enhance the earlier modeling process. In particular, the current frame, the last observable frame in the online video stream, serves as the direct visual cue for ongoing action recognition and the informative context for future action anticipation. As modeling errors accumulate over time, the resulting representations may progressively deviate from the actual semantics. Findings in cognitive neuroscience show that the hippocampus performs backward replay after observation to reinforce and correct the interpretation of previous observations. Inspired by this, we propose to incorporate backward temporal modeling following forward temporal modeling, enabling the model to leverage backward temporal modeling to enhance forward temporal modeling. Based on this idea, we propose a unified model for OAD and OAA, named Bidirectional Online Mamba (BiOMamba). Specifically, to address the excessive length and relevance imbalance in observable sequences, BiOMamba compresses distant long-term memory and preserves recent short-term memory. Then, BiOMamba sequentially model both forward and backward temporal dependencies in the whole memory. Finally, according to the temporal modeling result, BiOMamba generates representations for current and future actions. BiOMamba achieves state-of-the-art performance on THUMOS'14 (OAD: 73.3% mAP, OAA: 59.7% mAP) and TVSeries (OAD: 89.9% mcAP, OAA: 83.7% mcAP).
Lane topology reasoning plays a critical role in autonomous driving by modeling the connections among lanes and the topological relationships between lanes and traffic elements. Most existing methods adopt a first-detect-then-reason paradigm, where topological relationships are supervised based on the one-to-one assignment results obtained during the detection stage. This supervision strategy results in suboptimal topology reasoning performance due to the limited range of valid supervision. In this paper, we propose RATopo, a Redundancy Assignment strategy for lane Topology reasoning that enables quantity-rich and geometry-diverse topology supervision. Specifically, we restructure the Transformer decoder by swapping the cross-attention and self-attention layers. This allows redundant lane predictions to be retained before suppression, enabling effective one-to-many assignment. We also instantiate multiple parallel cross-attention blocks with independent parameters, which further enhances the diversity of detected lanes. Extensive experiments on OpenLane-V2 demonstrate that our RATopo strategy is model-agnostic and can be seamlessly integrated into existing topology reasoning frameworks, consistently improving both lane-lane and lane-traffic topology performance (e.g., + 15.7% and + 9.5% on TopoLogic for TOP_ll and TOP_lt on OpenLane-V2 subset_B, respectively).
Deep graph clustering, which aims to uncover the underlying structure within graphs and partition nodes into distinct groups, is a challenging research spot. However, the formation of the cluster in real-world graphs typically governed by the highly complex interaction of many underlying latent factors. Existing methods typically rely on the features and structure associated with the graph, and neglect the entanglement of these factors, resulting in sub-optimal clustering performance. In this paper, we propose a novel deep graph clustering framework named DisenCluster, which learns disentangled representations to simultaneously consider node separation results from diverse perspectives. Specifically, we introduce a disentangled graph encoder that iteratively identifies the latent factors of the input graph by modeling the distribution over different factors for each edge. Subsequently, we utilize a factor-wise contrastive loss to encourage clustering-friendly disentangled representations, allowing us to derive different clustering results based on the corresponding factor. These results are then structured as anchor graphs and seamlessly integrated into a unified graph. Finally, we formulate the framework as a continuous relaxation of the high-order graph cut problem and optimize the objective to obtain effective cluster assignments. Results from experiments on a variety of publicly available datasets further reveal the effectiveness and superiority of our DisenCluster compared with baselines.
The spread of tampered text poses a critical challenge to information security. Previous methods for tampered text detection (TTD) primarily relied on visual artifacts as clues, while overlooking potential semantic inconsistencies introduced during text manipulation. To address this limitation, we propose TVSIP (Tampered text Visual-Semantic InterPreter), a novel framework leveraging Multimodal Large Language Models (MLLMs) to integrate both visual and semantic clues for comprehensive tampered text analysis and verification. TVSIP consists of a Locator and an Interpreter. The Locator combines the visual detection ability of existing expert models with the semantic comprehension capabilities of MLLMs to create precise tampering masks. Subsequently, the Interpreter provides comprehensive descriptions and explanations based on identified tampered regions. To train and evaluate TVSIP, we construct the TextDDLE benchmark using GPT-4o. Extensive experiments demonstrate that TVSIP outperforms expert models in pixel-level localization and advanced MLLMs in interpretability. Furthermore, it maintains robustness against image degradation and exhibits strong generalization ability on out-of-domain datasets. Our work highlights the crucial role of semantic inconsistencies in TTD and establishes a more reliable verification system for ensuring document authenticity in the digital age.
Large Language Models (LLMs) are trained on a vast amount of procedural texts, but they do not directly observe real-world phenomena. In the context of cooking recipes, this poses a challenge, as intermediate states of ingredients are often omitted, making it difficult for models to track ingredient states and understand recipes accurately. In this paper, we apply state probing, a method for evaluating a language model's understanding of the world, to the domain of cooking. We propose a new task and dataset for evaluating how well LLMs can recognize intermediate ingredient states during cooking procedures. We first construct a new Japanese recipe dataset with clear and accurate annotations of ingredient state changes, collected from well-structured and controlled recipe texts. Using this dataset, we design three novel tasks to evaluate whether LLMs can track ingredient state transitions and identify ingredients present at intermediate steps. Our experiments with widely used LLMs, such as Llama3.1-70B and Qwen2.5-72B, show that learning ingredient state knowledge improves their understanding of cooking processes, achieving performance comparable to commercial LLMs. The dataset are publicly available at: https://huggingface.co/datasets/mashi6n/nhkrecipe-100-anno-1.