CVPR 2025 Papers — Page 26

Task-Agnostic Guided Feature Expansion for Class-Incremental Learning

Bowen Zheng (Nanjing University), De-Chuan Zhan (Nanjing University)

🎯 What it does: A framework named TagFex is proposed to address the feature collision problem in class-incremental learning by continuously capturing task-irrelevant features to help the model learn new tasks while retaining old knowledge.

Task-Aware Clustering for Prompting Vision-Language Models

Fusheng Hao (Shenzhen Institute of Advanced Technology), Jun Cheng (Shenzhen Institute of Advanced Technology)

🎯 What it does: A Task-Aware Clustering (TAC) framework is proposed to enhance the prompt learning effectiveness of visual language models by generating lightweight task-aware pre-contexts.

Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding

Wenbo Chen (South China University of Technology), Hau-San Wong (City University of Hong Kong)

🎯 What it does: A task-aware cross-modal feature refinement transformer (TCRT) is proposed, which integrates large language models (LLM) with visual features for visual localization and segmentation tasks.

Task-driven Image Fusion with Learnable Fusion Loss

Haowen Bai (Xi'an Jiaotong University), Shuang Xu (Northwestern Polytechnical University)

🎯 What it does: A task-driven learnable fusion loss framework (TDFusion) has been designed and implemented to achieve adaptive fusion of multimodal images.

Task-Specific Gradient Adaptation for Few-Shot One-Class Classification

Yunlong Li (Beijing Institute of Technology), Yuchen Ren (Beijing Institute of Technology)

🎯 What it does: A task-specific gradient adaptation method TSGA is proposed for few-shot single-class classification tasks.

Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting

Maochen Yang (Nanjing University), Yinghuan Shi (Nanjing University)

🎯 What it does: This paper proposes a semi-supervised crowd counting framework TMTB, which enhances data through diffusion-based image filling in background areas and captures global context using a visual state space model to achieve high-precision counting.

TASTE-Rob: Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation

Hongxiang Zhao (Chinese University of Hong Kong Shenzhen), Xiaoguang Han (Chinese University of Hong Kong Shenzhen)

🎯 What it does: A large-scale egocentric hand-object interaction video dataset, TASTE-Rob, was developed, and a three-stage pose refinement pipeline was proposed to generate high-quality task-oriented hand-object interaction videos, thereby enhancing the generalization ability of robot imitation learning.

Taxonomy-Aware Evaluation of Vision-Language Models

Vésteinn Snæbjarnarson (University of Copenhagen), Stella Frank (University of Copenhagen)

🎯 What it does: This paper proposes a hierarchical structure-based evaluation framework to measure the accuracy and specificity of open-ended text generated by visual language models (VLM) in fine-grained visual classification tasks.

TCFG: Tangential Damping Classifier-free Guidance

Mingi Kwon (Yonsei University), Youngjung Uh (Yonsei University)

🎯 What it does: A tangential damping unconditioned classifier guidance (TCFG) based on singular value decomposition is designed, which improves the sampling trajectory of CFG by removing the tangential components from the unconditioned scores to align them with the conditioned scores.

Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution

Zhiyuan You (Chinese University of Hong Kong), Chao Dong (Shenzhen Institutes of Advanced Technology)

🎯 What it does: A method for image quality assessment based on a multimodal large language model, DeQA-Score, is proposed, utilizing distributed soft labels and Fidelity loss to achieve precise score regression and distribution prediction.

Teller: Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation

Dingcheng Zhen (Shanghai Soulgate Technology Co.), Ming Tao (Shanghai Soulgate Technology Co.)

🎯 What it does: The first real-time, streaming audio-driven portrait animation autoregressive framework, Teller, is proposed, achieving smooth animations at up to 25 frames per second.

Temporal Action Detection Model Compression by Progressive Block Drop

Xiaoyong Chen (Shenzhen MSU-BIT University), Xiping Hu (Beijing Institute of Technology)

🎯 What it does: A progressive block drop method is proposed to compress the depth of temporal action detection models while keeping the width unchanged, achieving efficient inference.

Temporal Alignment-Free Video Matching for Few-shot Action Recognition

SuBeen Lee (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)

🎯 What it does: A time-alignment-free few-shot action recognition method called TEAM is designed and implemented, which provides a global representation of videos using a fixed number of pattern tokens without the need for frame or subsequence alignment.

Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts

Yu Cao (Queen Mary University of London), Shaogang Gong (Queen Mary University of London)

🎯 What it does: Analyzing the generation process of diffusion models, it is found that visual artifacts mainly occur during the mutation phase, and ASCED is proposed to achieve unsupervised artifact detection and real-time correction by monitoring the dynamics of anomaly scores.

Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks

Kairong Yu (Zhejiang University), Qi Xu (Dalian University of Technology)

🎯 What it does: This paper proposes a knowledge distillation method for spiking neural networks, utilizing temporal separation and entropy regularization to enhance the performance of the student network.

Temporally Consistent Object-Centric Learning by Contrasting Slots

Anna Manasyan (University of Tuebingen), Andrii Zadaianchuk (University of Amsterdam)

🎯 What it does: This paper proposes an unsupervised video object centering learning framework called Slot Contrast, aimed at enhancing the temporal consistency of object slots in videos.

TensoFlow: Tensorial Flow-based Sampler for Inverse Rendering

Chun Gu (Fudan University), Xiatian Zhu (University of Surrey)

🎯 What it does: The TensoFlow method is proposed to learn spatial and directional importance samplers in inverse rendering, improving the Monte Carlo estimation of the rendering equation.

Test-time Augmentation Improves Efficiency in Conformal Prediction

Divya Shanmugam (Massachusetts Institute of Technology), John Guttag (Massachusetts Institute of Technology)

🎯 What it does: Introduce Test-Time Augmentation (TTA) into the framework of conformal prediction, enhancing the efficiency of the prediction set by applying label-preserving transformations to the input and aggregating the results.

Test-Time Backdoor Detection for Object Detection Models

Hangtao Zhang (Huazhong University of Science and Technology), Leo Yu Zhang (Griffith University)

🎯 What it does: A black-box, general object detection model backdoor attack detection method called TRACE is proposed.

Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation

Xingguo Lv (Anhui University), Xuejun Li (Anhui University)

🎯 What it does: A testing domain transfer framework based on multi-graph matching is constructed, utilizing cosmic embedding to achieve cross-domain consistency and adaptive optimization in medical image segmentation.

Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability

Unki Park (Sungkyunkwan University), Jong Hwan Ko (Sungkyunkwan University)

🎯 What it does: A method for instance-specific image compression based on Test-Time Fine-Tuning (TTFT) is proposed, which adaptively fine-tunes the compression model using SVD-LoRA in the encoder and decoder to meet the needs of closed-set and open-set machine vision tasks.

Test-Time Visual In-Context Tuning

Jiahao Xie (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)

🎯 What it does: This paper proposes a Visual Context Tuning (VICT) method for test-time, which utilizes a single test sample to perform self-supervised fine-tuning on visual context learning models, thereby enhancing robustness under distribution shifts.

TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer

Jialun Liu (Baidu Inc), Jingdong Wang (Baidu Inc)

🎯 What it does: A method called TexGarment is proposed for efficiently generating consistent clothing textures in UV space, addressing the shortcomings of traditional methods in 3D consistency and texture diversity.

TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting

Bojun Xiong (Peking University), Zhouhui Lian (Peking University)

🎯 What it does: Using octrees to sample 3D mesh surfaces, placing 3D Gaussians at the centers of the finest leaf nodes, and employing 3D U-Net to regress Gaussian parameters, high-quality PBR materials are generated and can be directly baked into UV space;

Text Augmented Correlation Transformer For Few-shot Classification & Segmentation

Srinivasa Rao Nandam (Surrey Institute for People-Centred AI), Muhammad Awais (Surrey Institute for People-Centred AI)

🎯 What it does: A multi-modal few-shot classification and segmentation framework TA-CST that integrates text and images is proposed, capable of achieving FS-CS using only text support, and can also enhance performance through text augmentation.

Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps

Jeeyung Kim (Purdue University), Qiang Qiu (Purdue University)

🎯 What it does: A method is proposed to optimize potential noise during inference by minimizing the distance between the cross-attention similarity matrix and the text self-attention matrix, thereby improving the semantic alignment from text to image.

Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction

Shanshan Huang (Chongqing University), Li Liu (Peking University)

🎯 What it does: A text-based multi-concept fashion image editing framework T-FIT is proposed, which can achieve precise single-concept and multi-concept editing on a single image.

Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding

Wenxuan Guo (Tsinghua University), Jiwen Lu (Tsinghua University)

🎯 What it does: A single-stage sparse convolution 3D visual localization framework TSP3D is designed, achieving efficient 3D visual localization through text-guided sparse voxel pruning and completion.

Textured Gaussians for Enhanced 3D Scene Appearance Modeling

Brian Chao (Stanford University), Changil Kim (Meta)

🎯 What it does: Extend the 3D Gaussian Splatting model to allow each Gaussian to carry local textures and an alpha channel, achieving a representation of spatially varying colors and opacities.

TFCustom: Customized Image Generation with Time-Aware Frequency Feature Guidance

Mushui Liu (Zhejiang University), Siming Fu (Zhejiang University)

🎯 What it does: This paper proposes the TFCustom framework for zero-shot individual-driven image generation, achieving precise injection of reference image features through synchronized ReferenceNet, temporal frequency feature refinement, and reward loss, enhancing the details and consistency of generated images.

The Art of Deception: Color Visual Illusions and Diffusion Models

Alexandra Gomez-Villa (Computer Vision Center), Joost van den Weijer

🎯 What it does: The study investigates how the intermediate representations in the reverse process of diffusion models produce brightness/color biases similar to human visual illusions, and utilizes this phenomenon to generate new visual illusions.

The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf

Yanis Benidir, Clement Mallet (University Gustave Eiffel)

🎯 What it does: This paper proposes the HySCDG generation pipeline, which utilizes Stable Diffusion and ControlNet to achieve semantically guided inpainting on VHR images, creating a dual-temporal semantic change detection dataset FSC-180k that includes both real and rewritten images.

The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation

Yuhan Liu (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)

🎯 What it does: This paper studies cross-domain few-shot semantic segmentation (CDFSS), analyzing and addressing the phenomenon that negatively impacts target domain performance during source domain training.

The Devil is in Temporal Token: High Quality Video Reasoning Segmentation

Sitong Gong (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)

The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video Generation

Bingjie Gao (Shanghai Jiao Tong University), Yaohui Wang (Shanghai Artificial Intelligence Laboratory)

🎯 What it does: A retrieval-enhanced prompt optimization framework (RAPO) has been developed to improve the generation quality of text-to-video generation models.

The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models

Naveen George (Indian Institute of Technology), Konda Reddy Mopuri (Indian Institute of Technology)

🎯 What it does: The paper studies the machine unlearning methods for text-to-image diffusion models and reveals their instability, indicating that the concept of unlearning re-emerges after fine-tuning.

The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation

Marcus Nordström (KTH Royal Institute of Technology), Henrik Hult (KTH Royal Institute of Technology)

🎯 What it does: This paper proposes an approximate volume-preserving Gaussian field noise model and analyzes the optimal solutions and threshold effects of cross-entropy and soft Dice loss under label noise.

The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion

Changan Chen (Stanford University), Ehsan Adeli (Stanford University)

🎯 What it does: This work proposes a unified multimodal language model for generating 3D human actions from various inputs such as speech, text, or motion, and can simultaneously understand and generate movements.

The PanAf-FGBG Dataset: Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition

Otto Brookes (University of Bristol), Tilo Burghardt (Max Planck Institute for Evolutionary Anthropology)

🎯 What it does: Proposed the PanAf-FGBG dataset and studied the impact of background information on wild chimpanzee behavior recognition.

The Photographer's Eye: Teaching Multimodal Large Language Models to See, and Critique Like Photographers

Daiqing Qi (University of Virginia), Sheng Li (University of Virginia)

🎯 What it does: A multimodal large language model called PhotoEye, oriented towards the photographer's perspective, has been proposed, along with the construction of a large-scale professional aesthetic dataset called PhotoCritique and a benchmark called PhotoBench, enhancing the understanding of image aesthetics.

The Power of Context: How Multimodality Improves Image Super-Resolution

Kangfu Mei (Google), Mauricio Delbracio (Johns Hopkins University)

🎯 What it does: A multi-modal diffusion model MMSR is proposed, which uses various auxiliary modalities (text descriptions, depth maps, semantic segmentation, edges) to guide image super-resolution generation.

The Scene Language: Representing Scenes with Programs, Words, and Embeddings

Yunzhi Zhang (Stanford University), Jiajun Wu (Stanford University)

🎯 What it does: This paper proposes Scene Language, which jointly expresses 3D scenes through three modalities: programming, vocabulary, and neural embeddings, supporting inference, rendering, and editing from text or images.

Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems

Song Xia (Nanyang Technological University), Xudong Jiang (Nanyang Technological University)

🎯 What it does: A CEM algorithm based on conditional entropy maximization is proposed, which enhances the robustness against model inversion attacks by performing Gaussian mixture estimation on intermediate features in collaborative reasoning, and can be seamlessly integrated into existing redundancy defense methods.

Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels

Quanjiang Li (National University of Defense Technology), Jiahui Liao (National University of Defense Technology)

🎯 What it does: A deep multi-view multi-label learning framework based on information bottleneck theory is proposed, which can simultaneously handle view missingness and label noise.

Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields

Navami Kairanda (Max Planck Institute for Informatics), Vladislav Golyanik (Max Planck Institute for Informatics)

🎯 What it does: This paper proposes a monocular non-rigid 3D surface tracking method using continuous neural deformation fields and thin-shell physical priors.

Think Small, Act Big: Primitive Prompt Learning for Lifelong Robot Manipulation

Yuanqi Yao (Shanghai AI Laboratory), Dong Wang (Shanghai AI Laboratory)

🎯 What it does: A lifelong robotic manipulation framework based on Primitive Prompt Learning (PPL) is proposed, achieving multi-task knowledge transfer through two-stage pre-training and lifelong learning.

Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces

Jihan Yang (New York University), Saining Xie (New York University)

🎯 What it does: This paper proposes a video-based visual spatial intelligence benchmark, VSI-Bench, and systematically evaluates the spatial reasoning capabilities of multimodal large language models.

Three Cars Approaching within 100m! Enhancing Distant Geometry by Tri-Axis Voxel Scanning for Camera-based Semantic Scene Completion

Jongseong Bae (Yonsei University), Ha Young Kim (Yonsei University)

🎯 What it does: A novel camera-based 3D semantic scene completion model named ScanSSC is designed, utilizing a Scan Module and Scan Loss to enhance the reconstruction quality of distant geometry.

Three-view Focal Length Recovery From Homographies

Yaqing Ding (Czech Technical University in Prague), Zuzana Kukelova (Czech Technical University in Prague)

🎯 What it does: This paper proposes an efficient method to recover the camera focal length from the four-point correspondence observed by a three-view coplanar camera using single images from three different views.

Through-The-Mask: Mask-based Motion Trajectories for Image-to-Video Generation

Guy Yariv (Meta), Adam Polyak (Meta)

🎯 What it does: This paper proposes a two-stage image-to-video (I2V) generation framework called THROUGH-THE-MASK, which first generates mask-based motion trajectories and then uses them along with the original image and text prompts to generate complete videos, enabling coherent action generation for multiple objects.

TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction

Aishwarya Agarwal (International Institute of Information Technology Hyderabad), Vineet Gandhi (International Institute of Information Technology Hyderabad)

🎯 What it does: This paper addresses the single-source domain generalization (SSDG) problem and proposes a novel training framework called TIDE, which enables the model to learn local concepts and perform corrections during inference to enhance robustness against semantic domain shifts.

Tightening Robustness Verification of MaxPool-based Neural Networks via Minimizing the Over-Approximation Zone

Yuan Xiao (Nanjing University), Zhenyu Chen (Nanjing University)

🎯 What it does: For convolutional neural networks that include MaxPool, a verifier called Ti-Lin is proposed, which achieves robustness verification by constructing the tightest linear bounds at the neuron level.

Tiled Diffusion

Or Madar (Reichman University), Ohad Fried (Reichman University)

🎯 What it does: A method called Tiled Diffusion is proposed, which utilizes diffusion models to generate seamlessly stitchable images in various stitching scenarios such as self-stitching, one-to-one, and many-to-many.

Time of the Flight of the Gaussians: Optimizing Depth Indirectly in Dynamic Radiance Fields

Runfeng Li (Brown University), James Tompkin (Carnegie Mellon University)

🎯 What it does: A method is proposed for fast reconstruction of dynamic scenes captured by a monocular continuous wave time-of-flight (C-ToF) camera using 3D Gaussian splatting, directly optimizing the raw intensity quadruples, enabling the reconstruction of the geometry and motion of fast-moving objects with a single camera.

Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model

Feng Liu (University of Chinese Academy of Sciences), Fang Wan (University of Chinese Academy of Sciences)

🎯 What it does: To address the inference speed bottleneck of video diffusion models, a training-free caching strategy called TeaCache is proposed. It utilizes temporal embedding modulation of noise input to estimate the output difference of the model, thereby dynamically deciding whether to reuse cached outputs during inference, significantly improving speed.

TimeTracker: Event-based Continuous Point Tracking for Video Frame Interpolation with Non-linear Motion

Haoyue Liu (Huazhong University of Science and Technology), Luxin Yan

🎯 What it does: A continuous point tracking framework called TimeTracker based on event cameras is designed for high frame rate video frame interpolation, generating intermediate frames at arbitrary times through scene segmentation, trajectory tracking, global optimization, and frame refinement.

TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation

Yabiao Wang (Zhejiang University), Yong Liu (Zhejiang University)

🎯 What it does: The TIMotion framework is proposed to achieve text-driven human-to-human interaction action generation.

TinyFusion: Diffusion Transformers Learned Shallow

Gongfan Fang (National University of Singapore), Xinchao Wang (National University of Singapore)

🎯 What it does: By learning a differentiable deep pruning method, a lightweight diffusion Transformer is constructed while maintaining high recoverability.

TKG-DM: Training-free Chroma Key Content Generation Diffusion Model

Ryugo Morita (Hosei University), Jinjia Zhou (DFKI GmbH)

🎯 What it does: Without the need for any fine-tuning, foreground objects are generated using diffusion models and synthesized on a background with a specified color key, achieving controllable separation of foreground and background;

Token Cropr: Faster ViTs for Quite a Few Tasks

Benjamin Bergner (Hasso Plattner Institute for Digital Engineering), Aravindh Mahendran (Google DeepMind)

🎯 What it does: We propose Cropr, a module that learns task relevance and efficiently prunes in Vision Transformers through cross-attention routing and auxiliary prediction heads, applicable to tasks such as classification, semantic segmentation, and object detection.

TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation

Liao Qu (ByteDance), Xinglong Wu (ByteDance)

🎯 What it does: Proposes TokenFlow, a unified image tokenizer that can support both multimodal understanding and image generation;

TokenHSI: Unified Synthesis of Physical Human-Scene Interactions through Task Tokenization

Liang Pan (Shanghai AI Laboratory), Jingbo Wang (Shanghai AI Laboratory)

🎯 What it does: This paper presents TokenHSI, a transformer-based unified controller that enables the synthesis of various physically feasible human-scene interactions (such as following, sitting, climbing, and carrying).

Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images

Jiuchen Chen (Beijing Institute of Technology), Kaiqi Li (Beijing Institute of Technology)

🎯 What it does: A dehazing method called DehazeXL has been developed, capable of end-to-end processing of ultra-high-resolution images on a single GPU, directly dehazing images of 8192×8192 and even 10240×10240 pixels.

TokenMotion: Decoupled Motion Control via Token Disentanglement for Human-centric Video Generation

Ruineng Li (OPPO US AI Center), Chiuman Ho (OPPO US AI Center)

🎯 What it does: The TokenMotion framework is proposed, achieving joint control of camera motion and human actions in video generation, supporting both text-to-video and image-to-video generation paradigms.

TopNet: Transformer-Efficient Occupancy Prediction Network for Octree-Structured Point Cloud Geometry Compression

Xinjie Wang (National University of Defense Technology), Hanyun Wang (Information Engineering University)

🎯 What it does: A Transformer-efficient occupancy prediction network called TopNet is proposed for octree point cloud geometry compression.

TopoCellGen: Generating Histopathology Cell Topology with a Diffusion Model

Meilong Xu (Stony Brook University), Chao Chen (Stony Brook University)

🎯 What it does: This paper proposes a cell topology generation method based on diffusion models, called TopoCellGen, to synthesize multi-class cell layouts while preserving realistic spatial and topological features.

TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model

Cheng Yang (Rutgers University), Bo Yuan (Rutgers University)

🎯 What it does: This paper proposes the TopV method, which performs one-time pruning of visual tokens in visual language models, significantly reducing computational and memory consumption during inference.

Tora: Trajectory-oriented Diffusion Transformer for Video Generation

Zhenghao Zhang (Alibaba Cloud Computing), Weizhi Wang (Alibaba Cloud Computing)

🎯 What it does: The Tora framework is proposed to achieve controllable video generation based on trajectories, text, and images, maintaining motion consistency across various resolutions, frame rates, and long durations (up to 204 frames).

Touch2Shape: Touch-Conditioned 3D Diffusion for Shape Exploration and Reconstruction

Yuanbo Wang (Dalian University of Technology), Xin Yang (Dalian University of Technology)

🎯 What it does: This paper proposes a tactile condition-based 3D diffusion model called Touch2Shape, which is used to detect and reconstruct the 3D shape of objects through continuous touch.

Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption

Du Chen (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)

🎯 What it does: A general full-reference image quality assessment model A-FINE is proposed, capable of evaluating image quality under the condition of imperfect reference image quality; a new database DiffIQA (180k images enhanced using diffusion models) and SRIQA-Bench (10 super-resolution results for 100 images) are constructed for training and evaluation.

Toward Real-world BEV Perception: Depth Uncertainty Estimation via Gaussian Splatting

Shu-Wei Lu (Atmanity Inc.), Yi-Ting Chen (National Yang Ming Chiao Tung University)

🎯 What it does: In this study, the authors propose the GaussianLSS framework, which models depth uncertainty by modeling the depth distribution for each pixel and converting it into a 3D Gaussian distribution. They then use the Gaussian Splatting method to generate efficient BEV features in a bird's-eye view.

Toward Robust Neural Reconstruction from Sparse Point Sets

Amine Ouasfi (Inria), Adnane Boukhayma (Inria)

🎯 What it does: This paper studies the method of unsupervised learning of signature distance functions under sparse noisy point clouds.

Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content

Rohit Kundu (Google), Amit K. Roy-Chowdhury (University of California)

🎯 What it does: A general video forgery detection model called UNITE is proposed, capable of simultaneously identifying facial manipulation, background manipulation, and fully AI-generated videos.

Towards All-in-One Medical Image Re-Identification

Yuan Tian (Shanghai AI Laboratory), Guangtao Zhai (Shanghai Jiao Tong University)

🎯 What it does: A unified multimodal medical image re-identification (MedReID) model called MaMI is proposed, which achieves the adaptation of multiple medical image modalities through a Continuous Modal Parameter Adapter (ComPA); at the same time, it utilizes a Medical Foundation Model (MFM) for medical prior alignment to enhance the discriminative power of features.

Towards Autonomous Micromobility through Scalable Urban Simulation

Wayne Wu (University of California), Bolei Zhou (University of California)

🎯 What it does: This paper presents a scalable urban simulation platform called URBAN-SIM, and based on this platform, constructs the URBAN-BENCH task set to evaluate the autonomy of various micro mobile robots in real urban environments.

Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards

Zijing Hu (Zhejiang University), Wenwu Zhu (Tsinghua University)

🎯 What it does: Fine-tuning a text-to-image diffusion model using reinforcement learning to address the issue of insufficient alignment between generated images and text prompts.

Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters

Xiaohan Qin (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)

🎯 What it does: This paper studies the gradient conflict problem in multi-task learning and proposes ConsMTL, which consistently alleviates conflicts by simultaneously updating shared parameters and task-specific parameters through a dual-layer optimization.

Towards Continual Universal Segmentation

Zihan Lin (University of Science and Technology of China), Xu Wang (University of Science and Technology of China)

🎯 What it does: Proposes the CUE framework, which breaks down the continuous segmentation task into two steps: high-level semantic understanding and low-level refinement, achieving a balance between stability and plasticity in continuous learning.

Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning

Tianxiang Yin (Nanjing University of Aeronautics and Astronautics), Han Sun (Nanjing University of Aeronautics and Astronautics)

🎯 What it does: This paper proposes a method that combines semi-supervised learning (SSL) and active learning (AL) to reduce labeling costs. By incorporating the AL objective into the mainstream pseudo-label-based SSL framework, the key differences between SSAL and traditional AL scenarios are clarified, and a feature re-alignment module is proposed to improve algorithm performance.

Towards Effective and Sparse Adversarial Attack on Spiking Neural Networks via Breaking Invisible Surrogate Gradients

Li Lun (Peking University), Xiaoxin Cui (Peking University)

🎯 What it does: Proposed potential-dependent pseudo-gradient (PDSG) and sparse dynamic attack (SDA) to generate high success rate adversarial samples on inferable spiking neural networks.

Towards Efficient Foundation Model for Zero-shot Amodal Segmentation

Zhaochen Liu (Peking University), Tingting Jiang (Meituan Inc.)

🎯 What it does: The SAMBA model is proposed to achieve zero-shot occlusion segmentation, utilizing SAM pre-trained knowledge for segmentation, providing an efficient and stable baseline.

Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency

Yikai Wang (Fudan University), Yanwei Fu (Fudan University)

🎯 What it does: The ASUKA framework is proposed, which utilizes post-training techniques to significantly reduce the insertion of irrelevant objects and color inconsistencies while maintaining the generative capabilities of the original diffusion/flow model, thereby improving image filling quality.

Towards Explainable and Unprecedented Accuracy in Matching Challenging Finger Crease Patterns

Zhenyu Zhou (Hong Kong Polytechnic University), Ajay Kumar (Hong Kong Polytechnic University)

🎯 What it does: A framework for interpretable finger joint texture matching based on keypoint similarity graph neural networks has been designed, and a multi-pose contactless finger joint video database with 351 subjects and 805,768 images has been made publicly available.

Towards Explicit Geometry-Reflectance Collaboration for Generalized LiDAR Segmentation in Adverse Weather

Longyu Yang (University of Electronic Science and Technology of China), Xiaofeng Zhu (University of Electronic Science and Technology of China)

🎯 What it does: This paper proposes the Geometry-Reflectance Collaboration (GRC) framework, specifically addressing the domain generalization problem of LiDAR semantic segmentation under adverse weather conditions. It uses a dual-branch approach to independently encode geometric structures and reflectance intensity, achieving robust fusion through information constraints and multi-layer feature collaboration.

Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition

Lintong Zhang (Korea University), Seong-Whan Lee (Korea University)

🎯 What it does: A Fine-grained Visual Counterfactual Explanation (FG-VCE) framework is proposed to generate object-level and part-level fine-grained contrastive explanations for misclassified samples, helping humans understand model decisions.

Towards General Visual-Linguistic Face Forgery Detection

Ke Sun (Xiamen University), Rongrong Ji (Xiamen University)

🎯 What it does: This study proposes a Facial Forgery Text Generator (FFTG) that utilizes masks and handcrafted features to generate original text annotations. It then refines these annotations through structured prompts to enhance GPT-4o-mini, followed by fine-tuning CLIP and multimodal LLMs with these high-quality texts to improve the performance and interpretability of facial forgery detection.

Towards Generalizable Scene Change Detection

Jae-Woo Kim (Gwangju Institute of Science and Technology), Ue-Hwan Kim (Gwangju Institute of Science and Technology)

🎯 What it does: A zero-shot, generalized scene change detection framework called GeSCF is proposed, which can accurately extract change areas under different time points and environments.

Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting

Kaouther Messaoud (École Polytechnique Fédérale de Lausanne), Alexandre Alahi (École Polytechnique Fédérale de Lausanne)

🎯 What it does: Proposes the PerReg+ framework, utilizing Perceiver IO to achieve dual-layer representation learning, registration queries, and adaptive prompt tuning, enhancing the generalization and multimodal capabilities of vehicle trajectory prediction.

Towards High-fidelity 3D Talking Avatar with Personalized Dynamic Texture

Xuanchen Li (Shanghai Jiao Tong University), Yichao Yan (Shanghai Jiao Tong University)

🎯 What it does: The TexTalk4D dataset and TexTalker model are proposed, achieving speech-driven 3D facial animation and synchronously generating high-quality dynamic textures.

Towards Human-Understandable Multi-Dimensional Concept Discovery

Arne Grobrügge (Karlsruhe Institute of Technology), Philipp Spitzer (Karlsruhe Institute of Technology)

🎯 What it does: A new interpretable method HU-MCD is proposed, which automatically discovers human-understandable concepts and provides reliable concept importance scores.

Towards Improved Text-Aligned Codebook Learning: Multi-Hierarchical Codebook-Text Alignment with Long Text

Guotao Liang (Harbin Institute of Technology), Yao He (SiFar Company)

🎯 What it does: A new text enhancement codebook learning framework TA-VQ is proposed, which generates longer texts through a visual language model to improve text alignment in codebook learning.

Towards In-the-wild 3D Plane Reconstruction from a Single Image

Jiachen Liu (Pennsylvania State University), Hengkai Guo (Bytedance)

🎯 What it does: The ZeroPlane framework is proposed to achieve 3D plane reconstruction from a single image across indoor and outdoor environments.

Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method

Xinshuai Song (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)

🎯 What it does: This paper proposes the Long-Horizon Visual Language Navigation (LH-VLN) task, constructs an automated bidirectional multi-granularity data generation platform NavGen and the LHPR-VLN benchmark, and introduces a Multi-Granularity Dynamic Memory (MGDM) model to address the issues of long task planning and memory integration.

Towards Lossless Implicit Neural Representation via Bit Plane Decomposition

Woo Kyoung Han (Korea University), Kyong Hwan Jin (Korea University)

🎯 What it does: This study reduces the parameter upper limit of implicit neural networks (INR) through a bit-plane decomposition method, achieving lossless reconstruction of 16-bit images and audio signals.

Towards Million-Scale Adversarial Robustness Evaluation With Stronger Individual Attacks

Yong Xie (Fudan University), Xingjun Ma (Fudan University)

🎯 What it does: A new single-instance attack method called Probability Margin Attack (PMA) is proposed, which enhances the effectiveness of adversarial robustness assessment through the Probability Margin Loss in probability space.

Towards More General Video-based Deepfake Detection through Facial Component Guided Adaptation for Foundation Model

Yue-Hua Han (Academia Sinica), Jun-Cheng Chen (Academia Sinica)

🎯 What it does: A side network decoder based on the CLIP image encoder is proposed, which includes spatial and temporal modules, and detects deepfake videos through Face Component Guidance (FCG) loss.

Towards Natural Language-Based Document Image Retrieval: New Dataset and Benchmark

Hao Guo (Institute of Information Engineering), Hailun Lin (Institute of Information Engineering)

🎯 What it does: This paper studies the task of document image retrieval based on natural language queries (NL-DIR), proposing a corresponding benchmark dataset and an efficient retrieval framework that combines two-stage recall and re-ranking.

Towards Open-Vocabulary Audio-Visual Event Localization

Jinxing Zhou (Mohamed bin Zayed University of Artificial Intelligence), Meng Wang (Hefei University of Technology)

🎯 What it does: This paper proposes the Open Vocabulary Audio-Video Event Localization (OV-AVEL) task and provides two baseline schemes: training with frozen parameters and fine-tuning.

Towards Optimizing Large-Scale Multi-Graph Matching in Bioimaging

Max Kahl (Heidelberg University), Bogdan Savchynskyy (Heidelberg University)

🎯 What it does: This paper proposes an efficient algorithm GREEDA for large-scale incomplete multi-image matching in biological imaging, addressing the infeasibility of traditional complete multi-image matching methods in sparse incomplete instances.

Towards Practical Real-Time Neural Video Compression

Zhaoyang Jia (University of Science and Technology of China), Yan Lu (Microsoft Research Asia)

🎯 What it does: A real-time neural video encoder (DCVC-RT) is proposed, achieving high compression ratios, low latency, and wide applicability.