arXiv subjects

Yuxin Peng

Publications and source records attributed to Yuxin Peng.

At least 37 records · Page 2Linked to original sources

arXiv2026

Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation

Recent advancements in Large Vision-Language Models (LVLMs) have demonstrated remarkable multimodal perception capabilities, garnering significant attention. While numerous evaluation studies have emerged, assessing LVLMs both holistically and on specialized tasks, fine-grained image tasks-fundamental to computer vision-remain largely unexplored. To fill this gap, we introduce a comprehensive fine-grained evaluation benchmark, i.e., FG-BMK, comprising 1.01 million questions and 0.33 million images. Our evaluation systematically examines LVLMs from both human-oriented and machine-oriented perspectives, focusing on their semantic recognition and fine-grained feature representation capabilities. Through extensive experiments on twelve representative LVLMs/VLMs, we uncover key findings regarding the influence of training paradigms, modality alignment, perturbation susceptibility, and fine-grained category reasoning on task performance. This work provides critical insights into the limitations of current LVLMs and offers guidance for future data construction and model design in the development of more advanced LVLMs. Our code is open-source and available at https://github.com/SEU-VIPGroup/FG-BMK.

cs.CV↗

arXiv2026

OmniVTG: A Large-Scale Dataset and Training Paradigm for Open-World Video Temporal Grounding

Video Temporal Grounding (VTG), the task of localizing video segments from text queries, struggles in open-world settings due to limited dataset scale and semantic diversity, causing performance gaps between common and rare concepts. To overcome these limitations, we introduce OmniVTG, a new large-scale dataset for open-world VTG, coupled with a Self-Correction Chain-of-Thought (CoT) training paradigm designed to enhance the grounding capabilities of Multimodal Large Language Models (MLLMs). Our OmniVTG is constructed via a novel Semantic Coverage Iterative Expansion pipeline, which first identifies gaps in the vocabulary of existing datasets and collects videos that are highly likely to contain these target concepts. For high-quality annotation, we leverage the insight that modern MLLMs excel at dense captioning more than direct grounding and design a caption-centric data engine to prompt MLLMs to generate dense, timestamped descriptions. Beyond the dataset, we observe that simple supervised finetuning (SFT) is insufficient, as a performance gap between rare and common concepts still persists. We find that MLLMs' video understanding ability significantly surpasses their direct grounding ability. Based on this, we propose a Self-Correction Chain-of-Thought (CoT) training paradigm. We train the MLLM to first predict, then use its understanding capabilities to reflect on and refine its own predictions. This capability is instilled via a three-stage pipeline of SFT, CoT finetuning, and reinforcement learning. Extensive experiments show our approach not only excels at open-world grounding in our OmniVTG dataset but also achieves state-of-the-art zero-shot performance on four existing VTG benchmarks. Code is available at https://github.com/oceanflowlab/OmniVTG.

cs.CV↗

arXiv2026

Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning

Any entity in the visual world can be hierarchically grouped based on shared characteristics and mapped to fine-grained sub-categories. While Multi-modal Large Language Models (MLLMs) achieve strong performance on coarse-grained visual tasks, they often struggle with Fine-Grained Visual Recognition (FGVR). Adapting general-purpose MLLMs to FGVR typically requires large amounts of annotated data, which is costly to obtain, leaving a substantial performance gap compared to contrastive CLIP models dedicated for discriminative tasks. Moreover, MLLMs tend to overfit to seen sub-categories and generalize poorly to unseen ones. To address these challenges, we propose Fine-R1, an MLLM tailored for FGVR through an R1-style training framework: (1) Chain-of-Thought Supervised Fine-tuning, where we construct a high-quality FGVR CoT dataset with rationales of "visual analysis, candidate sub-categories, comparison, and prediction", transition the model into a strong open-world classifier; and (2) Triplet Augmented Policy Optimization, where Intra-class Augmentation mixes trajectories from anchor and positive images within the same category to improve robustness to intra-class variance, while Inter-class Augmentation maximizes the response distinction conditioned on images across sub-categories to enhance discriminative ability. With only 4-shot training, Fine-R1 outperforms existing general MLLMs, reasoning MLLMs, and even contrastive CLIP models in identifying both seen and unseen sub-categories, showing promise in working in knowledge-intensive domains where gathering expert annotations for all sub-categories is arduous. Code is available at https://github.com/PKU-ICST-MIPL/FineR1_ICLR2026.

cs.CV↗

arXiv2026

Taxonomy-Aware Representation Alignment for Hierarchical Visual Recognition with Large Multimodal Models

A high-performing, general-purpose visual understanding model should map visual inputs to a taxonomic tree of labels, identify novel categories beyond the training set for which few or no publicly available images exist. Large Multimodal Models (LMMs) have achieved remarkable progress in fine-grained visual recognition (FGVR) for known categories. However, they remain limited in hierarchical visual recognition (HVR) that aims at predicting consistent label paths from coarse to fine categories, especially for novel categories. To tackle these challenges, we propose Taxonomy-Aware Representation Alignment (TARA), a simple yet effective strategy to inject taxonomic knowledge into LMMs. TARA leverages representations from biology foundation models (BFMs) that encode rich biological relationships through hierarchical contrastive learning. By aligning the intermediate representations of visual features with those of BFMs, LMMs are encouraged to extract discriminative visual cues well structured in the taxonomy tree. Additionally, we align the representations of the first answer token with the ground-truth label, flexibly bridging the gap between contextualized visual features and categories of varying granularity according to user intent. Experiments demonstrate that TARA consistently enhances LMMs' hierarchical consistency and leaf node accuracy, enabling reliable recognition of both known and novel categories within complex biological taxonomies. Code is available at https://github.com/PKU-ICST-MIPL/TARA_CVPR2026.

cs.CV↗

arXiv2026

Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models

Recently, unified multimodal models (UMMs) have made remarkable progress in integrating visual understanding and generation, demonstrating strong potential for complex text-to-image (T2I) tasks. Despite their theoretical promise, a persistent capability gap exists: UMMs typically exhibit superior visual understanding but comparatively weaker generative capabilities. This discrepancy arises largely from the intrinsic decoupling between the understanding and generation processes. While a UMM can accurately interpret fine-grained visual details, it often struggles to produce semantically coherent images from complex textual prompts. To address this challenge, we explore UMMs' internal understanding capability to enhance generation quality. We propose a token-level intrinsic text-image alignment reward mechanism, GvU, enabling the UMM to act simultaneously as teacher and student: it evaluates its own outputs using the understanding branch to guide the generations accordingly. Building upon this, we design a self-supervised reinforcement learning framework, allowing UMMs to iteratively improve their generation quality through understanding-based intrinsic reward signals--without reliance on external supervision. Experimental results show that our method substantially boosts UMMs' generation, which in turn strengthens their fine-grained visual understanding, narrowing the capability gap between UMMs' visual understanding and generation.

cs.CV↗

arXiv2026

Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping

The widespread use of smartphones has made photography ubiquitous, yet a clear gap remains between ordinary users and professional photographers, who can identify aesthetic issues and provide actionable shooting guidance during capture. We define this capability as aesthetic guidance (AG) -- an essential but largely underexplored domain in computational aesthetics. Existing multimodal large language models (MLLMs) primarily offer overly positive feedback, failing to identify issues or provide actionable guidance. Without AG capability, they cannot effectively identify distracting regions or optimize compositional balance, thus also struggling in aesthetic cropping, which aims to refine photo composition through reframing after capture. To address this, we introduce AesGuide, the first large-scale AG dataset and benchmark with 10,748 photos annotated with aesthetic scores, analyses, and guidance. Building upon it, we propose Venus, a two-stage framework that first empowers MLLMs with AG capability through progressively complex aesthetic questions and then activates their aesthetic cropping power via CoT-based rationales. Extensive experiments show that Venus substantially improves AG capability and achieves state-of-the-art (SOTA) performance in aesthetic cropping, enabling interpretable and interactive aesthetic refinement across both stages of photo creation. Code is available at https://github.com/PKU-ICST-MIPL/Venus_CVPR2026.

cs.CV↗

arXiv2026

TiFRe: Text-guided Video Frame Reduction for Efficient Video Multi-modal Large Language Models

With the rapid development of Large Language Models (LLMs), Video Multi-Modal Large Language Models (Video MLLMs) have achieved remarkable performance in video-language tasks such as video understanding and question answering. However, Video MLLMs face high computational costs, particularly in processing numerous video frames as input, which leads to significant attention computation overhead. A straightforward approach to reduce computational costs is to decrease the number of input video frames. However, simply selecting key frames at a fixed frame rate (FPS) often overlooks valuable information in non-key frames, resulting in notable performance degradation. To address this, we propose Text-guided Video Frame Reduction (TiFRe), a framework that reduces input frames while preserving essential video information. TiFRe uses a Text-guided Frame Sampling (TFS) strategy to select key frames based on user input, which is processed by an LLM to generate a CLIP-style prompt. Pre-trained CLIP encoders calculate the semantic similarity between the prompt and each frame, selecting the most relevant frames as key frames. To preserve video semantics, TiFRe employs a Frame Matching and Merging (FMM) mechanism, which integrates non-key frame information into the selected key frames, minimizing information loss. Experiments show that TiFRe effectively reduces computational costs while improving performance on video-language tasks.

cs.CV↗

arXiv2026

Multi-Resolution Alignment for Voxel Sparsity in Camera-Based 3D Semantic Scene Completion

Camera-based 3D semantic scene completion (SSC) offers a cost-effective solution for assessing the geometric occupancy and semantic labels of each voxel in the surrounding 3D scene with image inputs, providing a voxel-level scene perception foundation for the perception-prediction-planning autonomous driving systems. Although significant progress has been made in existing methods, their optimization rely solely on the supervision from voxel labels and face the challenge of voxel sparsity as a large portion of voxels in autonomous driving scenarios are empty, which limits both optimization efficiency and model performance. To address this issue, we propose a \textit{Multi-Resolution Alignment (MRA)} approach to mitigate voxel sparsity in camera-based 3D semantic scene completion, which exploits the scene and instance level alignment across multi-resolution 3D features as auxiliary supervision. Specifically, we first propose the Multi-resolution View Transformer module, which projects 2D image features into multi-resolution 3D features and aligns them at the scene level through fusing discriminative seed features. Furthermore, we design the Cubic Semantic Anisotropy module to identify the instance-level semantic significance of each voxel, accounting for the semantic differences of a specific voxel against its neighboring voxels within a cubic area. Finally, we devise a Critical Distribution Alignment module, which selects critical voxels as instance-level anchors with the guidance of cubic semantic anisotropy, and applies a circulated loss for auxiliary supervision on the critical feature distribution consistency across different resolutions. The code is available at https://github.com/PKU-ICST-MIPL/MRA_TIP.

cs.CV↗

arXiv2025

Bi-C2R: Bidirectional Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification

Lifelong person Re-IDentification (L-ReID) exploits sequentially collected data to continuously train and update a ReID model, focusing on the overall performance of all data. Its main challenge is to avoid the catastrophic forgetting problem of old knowledge while training on new data. Existing L-ReID methods typically re-extract new features for all historical gallery images for inference after each update, known as "re-indexing". However, historical gallery data typically suffers from direct saving due to the data privacy issue and the high re-indexing costs for large-scale gallery images. As a result, it inevitably leads to incompatible retrieval between query features extracted by the updated model and gallery features extracted by those before the update, greatly impairing the re-identification performance. To tackle the above issue, this paper focuses on a new task called Re-index Free Lifelong person Re-IDentification (RFL-ReID), which requires performing lifelong person re-identification without re-indexing historical gallery images. Therefore, RFL-ReID is more challenging than L-ReID, requiring continuous learning and balancing new and old knowledge in diverse streaming data, and making the features output by the new and old models compatible with each other. To this end, we propose a Bidirectional Continuous Compatible Representation (Bi-C2R) framework to continuously update the gallery features extracted by the old model to perform efficient L-ReID in a compatible manner. We verify our proposed Bi-C2R method through theoretical analysis and extensive experiments on multiple benchmarks, which demonstrate that the proposed method can achieve leading performance on both the introduced RFL-ReID task and the traditional L-ReID task.

cs.CV↗

arXiv2025

CKDA: Cross-modality Knowledge Disentanglement and Alignment for Visible-Infrared Lifelong Person Re-identification

Lifelong person Re-IDentification (LReID) aims to match the same person employing continuously collected individual data from different scenarios. To achieve continuous all-day person matching across day and night, Visible-Infrared Lifelong person Re-IDentification (VI-LReID) focuses on sequential training on data from visible and infrared modalities and pursues average performance over all data. To this end, existing methods typically exploit cross-modal knowledge distillation to alleviate the catastrophic forgetting of old knowledge. However, these methods ignore the mutual interference of modality-specific knowledge acquisition and modality-common knowledge anti-forgetting, where conflicting knowledge leads to collaborative forgetting. To address the above problems, this paper proposes a Cross-modality Knowledge Disentanglement and Alignment method, called CKDA, which explicitly separates and preserves modality-specific knowledge and modality-common knowledge in a balanced way. Specifically, a Modality-Common Prompting (MCP) module and a Modality-Specific Prompting (MSP) module are proposed to explicitly disentangle and purify discriminative information that coexists and is specific to different modalities, avoiding the mutual interference between both knowledge. In addition, a Cross-modal Knowledge Alignment (CKA) module is designed to further align the disentangled new knowledge with the old one in two mutually independent inter- and intra-modality feature spaces based on dual-modality prototypes in a balanced manner. Extensive experiments on four benchmark datasets verify the effectiveness and superiority of our CKDA against state-of-the-art methods. The source code of this paper is available at https://github.com/PKU-ICST-MIPL/CKDA-AAAI2026.

cs.CV↗

arXiv2025

HD$^2$-SSC: High-Dimension High-Density Semantic Scene Completion for Autonomous Driving

Camera-based 3D semantic scene completion (SSC) plays a crucial role in autonomous driving, enabling voxelized 3D scene understanding for effective scene perception and decision-making. Existing SSC methods have shown efficacy in improving 3D scene representations, but suffer from the inherent input-output dimension gap and annotation-reality density gap, where the 2D planner view from input images with sparse annotated labels leads to inferior prediction of real-world dense occupancy with a 3D stereoscopic view. In light of this, we propose the corresponding High-Dimension High-Density Semantic Scene Completion (HD$^2$-SSC) framework with expanded pixel semantics and refined voxel occupancies. To bridge the dimension gap, a High-dimension Semantic Decoupling module is designed to expand 2D image features along a pseudo third dimension, decoupling coarse pixel semantics from occlusions, and then identify focal regions with fine semantics to enrich image features. To mitigate the density gap, a High-density Occupancy Refinement module is devised with a "detect-and-refine" architecture to leverage contextual geometric and semantic structures for enhanced semantic density with the completion of missing voxels and correction of erroneous ones. Extensive experiments and analyses on the SemanticKITTI and SSCBench-KITTI-360 datasets validate the effectiveness of our HD$^2$-SSC framework.

cs.CV↗

arXiv2025

SPHERE: Semantic-PHysical Engaged REpresentation for 3D Semantic Scene Completion

Camera-based 3D Semantic Scene Completion (SSC) is a critical task in autonomous driving systems, assessing voxel-level geometry and semantics for holistic scene perception. While existing voxel-based and plane-based SSC methods have achieved considerable progress, they struggle to capture physical regularities for realistic geometric details. On the other hand, neural reconstruction methods like NeRF and 3DGS demonstrate superior physical awareness, but suffer from high computational cost and slow convergence when handling large-scale, complex autonomous driving scenes, leading to inferior semantic accuracy. To address these issues, we propose the Semantic-PHysical Engaged REpresentation (SPHERE) for camera-based SSC, which integrates voxel and Gaussian representations for joint exploitation of semantic and physical information. First, the Semantic-guided Gaussian Initialization (SGI) module leverages dual-branch 3D scene representations to locate focal voxels as anchors to guide efficient Gaussian initialization. Then, the Physical-aware Harmonics Enhancement (PHE) module incorporates semantic spherical harmonics to model physical-aware contextual details and promote semantic-geometry consistency through focal distribution alignment, generating SSC results with realistic details. Extensive experiments and analyses on the popular SemanticKITTI and SSCBench-KITTI-360 benchmarks validate the effectiveness of SPHERE. The code is available at https://github.com/PKU-ICST-MIPL/SPHERE_ACMMM2025.

cs.CV↗

arXiv2025

SphereDrag: Spherical Geometry-Aware Panoramic Image Editing

Image editing has made great progress on planar images, but panoramic image editing remains underexplored. Due to their spherical geometry and projection distortions, panoramic images present three key challenges: boundary discontinuity, trajectory deformation, and uneven pixel density. To tackle these issues, we propose SphereDrag, a novel panoramic editing framework utilizing spherical geometry knowledge for accurate and controllable editing. Specifically, adaptive reprojection (AR) uses adaptive spherical rotation to deal with discontinuity; great-circle trajectory adjustment (GCTA) tracks the movement trajectory more accurate; spherical search region tracking (SSRT) adaptively scales the search range based on spherical location to address uneven pixel density. Also, we construct PanoBench, a panoramic editing benchmark, including complex editing tasks involving multiple objects and diverse styles, which provides a standardized evaluation framework. Experiments show that SphereDrag gains a considerable improvement compared with existing methods in geometric consistency and image quality, achieving up to 10.5% relative improvement.

cs.CV↗

arXiv2025

Identity-Preserving Text-to-Video Generation via Training-Free Prompt, Image, and Guidance Enhancement

Identity-preserving text-to-video (IPT2V) generation creates videos faithful to both a reference subject image and a text prompt. While fine-tuning large pretrained video diffusion models on ID-matched data achieves state-of-the-art results on IPT2V, data scarcity and high tuning costs hinder broader improvement. We thus introduce a Training-Free Prompt, Image, and Guidance Enhancement (TPIGE) framework that bridges the semantic gap between the video description and the reference image and design sampling guidance that enhances identity preservation and video quality, achieving performance gains at minimal cost.Specifically, we first propose Face Aware Prompt Enhancement, using GPT-4o to enhance the text prompt with facial details derived from the reference image. We then propose Prompt Aware Reference Image Enhancement, leveraging an identity-preserving image generator to refine the reference image, rectifying conflicts with the text prompt. The above mutual refinement significantly improves input quality before video generation. Finally, we propose ID-Aware Spatiotemporal Guidance Enhancement, utilizing unified gradients to optimize identity preservation and video quality jointly during generation.Our method outperforms prior work and is validated by automatic and human evaluations on a 1000 video test set, winning first place in the ACM Multimedia 2025 Identity-Preserving Video Generation Challenge, demonstrating state-of-the-art performance and strong generality. The code is available at https://github.com/Andyplus1/IPT2V.git.

cs.CV↗

arXiv2025

Investigating Domain Gaps for Indoor 3D Object Detection

As a fundamental task for indoor scene understanding, 3D object detection has been extensively studied, and the accuracy on indoor point cloud data has been substantially improved. However, existing researches have been conducted on limited datasets, where the training and testing sets share the same distribution. In this paper, we consider the task of adapting indoor 3D object detectors from one dataset to another, presenting a comprehensive benchmark with ScanNet, SUN RGB-D and 3D Front datasets, as well as our newly proposed large-scale datasets ProcTHOR-OD and ProcFront generated by a 3D simulator. Since indoor point cloud datasets are collected and constructed in different ways, the object detectors are likely to overfit to specific factors within each dataset, such as point cloud quality, bounding box layout and instance features. We conduct experiments across datasets on different adaptation scenarios including synthetic-to-real adaptation, point cloud quality adaptation, layout adaptation and instance feature adaptation, analyzing the impact of different domain gaps on 3D object detectors. We also introduce several approaches to improve adaptation performances, providing baselines for domain adaptive indoor 3D object detection, hoping that future works may propose detectors with stronger generalization ability across domains. Our project homepage can be found in https://jeremyzhao1998.github.io/DAVoteNet-release/.

cs.CV↗

arXiv2025

Interact-Custom: Customized Human Object Interaction Image Generation

Compositional Customized Image Generation aims to customize multiple target concepts within generation content, which has gained attention for its wild application. Existing approaches mainly concentrate on the target entity's appearance preservation, while neglecting the fine-grained interaction control among target entities. To enable the model of such interaction control capability, we focus on human object interaction scenario and propose the task of Customized Human Object Interaction Image Generation(CHOI), which simultaneously requires identity preservation for target human object and the interaction semantic control between them. Two primary challenges exist for CHOI:(1)simultaneous identity preservation and interaction control demands require the model to decompose the human object into self-contained identity features and pose-oriented interaction features, while the current HOI image datasets fail to provide ideal samples for such feature-decomposed learning.(2)inappropriate spatial configuration between human and object may lead to the lack of desired interaction semantics. To tackle it, we first process a large-scale dataset, where each sample encompasses the same pair of human object involving different interactive poses. Then we design a two-stage model Interact-Custom, which firstly explicitly models the spatial configuration by generating a foreground mask depicting the interaction behavior, then under the guidance of this mask, we generate the target human object interacting while preserving their identities features. Furthermore, if the background image and the union location of where the target human object should appear are provided by users, Interact-Custom also provides the optional functionality to specify them, offering high content controllability. Extensive experiments on our tailored metrics for CHOI task demonstrate the effectiveness of our approach.

cs.CV↗

arXiv2025

Advancing 3D Scene Understanding with MV-ScanQA Multi-View Reasoning Evaluation and TripAlign Pre-training Dataset

The advancement of 3D vision-language (3D VL) learning is hindered by several limitations in existing 3D VL datasets: they rarely necessitate reasoning beyond a close range of objects in single viewpoint, and annotations often link instructions to single objects, missing richer contextual alignments between multiple objects. This significantly curtails the development of models capable of deep, multi-view 3D scene understanding over distant objects. To address these challenges, we introduce MV-ScanQA, a novel 3D question answering dataset where 68% of questions explicitly require integrating information from multiple views (compared to less than 7% in existing datasets), thereby rigorously testing multi-view compositional reasoning. To facilitate the training of models for such demanding scenarios, we present TripAlign dataset, a large-scale and low-cost 2D-3D-language pre-training corpus containing 1M <2D view, set of 3D objects, text> triplets that explicitly aligns groups of contextually related objects with text, providing richer, view-grounded multi-object multimodal alignment signals than previous single-object annotations. We further develop LEGO, a baseline method for the multi-view reasoning challenge in MV-ScanQA, transferring knowledge from pre-trained 2D LVLMs to 3D domain with TripAlign. Empirically, LEGO pre-trained on TripAlign achieves state-of-the-art performance not only on the proposed MV-ScanQA, but also on existing benchmarks for 3D dense captioning and question answering. Datasets and code are available at https://matthewdm0816.github.io/tripalign-mvscanqa.

cs.CV↗

arXiv2025

TRKT: Weakly Supervised Dynamic Scene Graph Generation with Temporal-enhanced Relation-aware Knowledge Transferring

Dynamic Scene Graph Generation (DSGG) aims to create a scene graph for each video frame by detecting objects and predicting their relationships. Weakly Supervised DSGG (WS-DSGG) reduces annotation workload by using an unlocalized scene graph from a single frame per video for training. Existing WS-DSGG methods depend on an off-the-shelf external object detector to generate pseudo labels for subsequent DSGG training. However, detectors trained on static, object-centric images struggle in dynamic, relation-aware scenarios required for DSGG, leading to inaccurate localization and low-confidence proposals. To address the challenges posed by external object detectors in WS-DSGG, we propose a Temporal-enhanced Relation-aware Knowledge Transferring (TRKT) method, which leverages knowledge to enhance detection in relation-aware dynamic scenarios. TRKT is built on two key components:(1)Relation-aware knowledge mining: we first employ object and relation class decoders that generate category-specific attention maps to highlight both object regions and interactive areas. Then we propose an Inter-frame Attention Augmentation strategy that exploits optical flow for neighboring frames to enhance the attention maps, making them motion-aware and robust to motion blur. This step yields relation- and motion-aware knowledge mining for WS-DSGG. (2) we introduce a Dual-stream Fusion Module that integrates category-specific attention maps into external detections to refine object localization and boost confidence scores for object proposals. Extensive experiments demonstrate that TRKT achieves state-of-the-art performance on Action Genome dataset. Our code is avaliable at https://github.com/XZPKU/TRKT.git.

cs.CV↗