ICCV 2025 Papers — Page 24
IEEE/CVF International Conference on Computer Vision · 2701 papers
TAG-WM: Tamper-Aware Generative Image Watermarking via Diffusion Inversion Sensitivity
Yuzhuo Chen (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
GenerationData SynthesisAnomaly DetectionDiffusion modelImageStochastic Differential Equation
🎯 What it does: A generative watermarking framework TAG-WM is proposed, which simultaneously embeds copyright watermarks and local positioning watermarks during the generation process of diffusion models, and utilizes diffusion inversion sensitivity to achieve damage detection and adaptive decoding.
Talking to DINO: Bridging Self-Supervised Vision Backbones with Language for Open-Vocabulary Segmentation
Luca Barsellotti (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
SegmentationTransformerContrastive LearningImageText
🎯 What it does: A method named Talk2DINO is proposed, which combines the fine-grained visual features of DINOv2 and the textual semantics of CLIP to achieve unsupervised open vocabulary segmentation without the need for fine-tuning the underlying model.
Taming Flow Matching with Unbalanced Optimal Transport into Fast Pansharpening
Zihan Cao (University of Electronic Science and Technology of China), Liang-Jian Deng (University of Electronic Science and Technology of China)
Image TranslationRestorationFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: Proposes the OTFM framework, which utilizes unbalanced optimal transport combined with flow matching to achieve high-quality image stitching in a single step.
Taming the Untamed: Graph-Based Knowledge Retrieval and Reasoning for MLLMs to Conquer the Unknown
Bowen Wang (Osaka University), Yuta Nakashima (Osaka University)
RetrievalGraph Neural NetworkLarge Language ModelPrompt EngineeringMultimodalityGraphBenchmark
🎯 What it does: A multimodal knowledge graph themed on 'Monster Hunter: World' (MH-MMKG) was constructed, and a benchmark test consisting of 238 questions was designed. An untrained multi-agent retriever was proposed to allow MLLM to automatically acquire relevant knowledge from the graph, and then enhance reasoning to answer questions.
TAPNext: Tracking Any Point (TAP) as Next Token Prediction
Artem Zholus (Chandar Research Lab), Ross Goroshin (Google DeepMind)
Object TrackingTransformerSupervised Fine-TuningVideo
🎯 What it does: This paper proposes an online point tracking model called TAPNext, which transforms the task of tracking any point (TAP) into next token prediction, utilizing an interleaved structure of SSM and ViT for end-to-end training without specific tracking biases.
TAR3D: Creating High-Quality 3D Assets via Next-Part Prediction
Xuying Zhang (Nankai University), Ming-Ming Cheng (Nankai University)
GenerationData SynthesisTransformerLarge Language ModelAuto EncoderMesh
🎯 What it does: This paper proposes an autoregressive 3D asset generation framework based on a three-plane quantized variational autoencoder (3D VQ-VAE) and GPT, named TAR3D. It first compresses 3D meshes into discrete three-plane representations, and then uses a pre-trained GPT to progressively generate geometric components in a 'next-part' manner, resulting in high-quality 3D models.
Target Bias Is All You Need: Zero-Shot Debiasing of Vision-Language Models with Bias Corpus
Taeuk Jang (Purdue University), Xiaoqian Wang (Purdue University)
ClassificationRetrievalTransformerVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: A zero-shot, task-agnostic debiasing method called ZSDebias is proposed, which decomposes CLIP image embeddings using biased text corpora and public visual datasets to obtain biased and neutral subspaces, aiming to reduce the bias of VLM under unlabeled conditions.
TARO: Timestep-Adaptive Representation Alignment with Onset-Aware Conditioning for Synchronized Video-to-Audio Synthesis
Tri Ton (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisTransformerAuto EncoderVideoMultimodalityAudio
🎯 What it does: The TARO framework is proposed to achieve high-fidelity, temporally consistent video-to-audio synthesis.
TARS: Traffic-Aware Radar Scene Flow Estimation
Jialong Wu (University of Wuppertal), Matthias Rottmann (Osnabruck University)
Object DetectionAutonomous DrivingRecurrent Neural NetworkOptical FlowPoint Cloud
🎯 What it does: A radar point cloud scene flow estimation framework TARS based on traffic-level information is proposed, achieving joint learning of detection and scene flow.
Task Vector Quantization for Memory-Efficient Model Merging
Youngeun Kim (Yale University), Sungeun Hong (Sungkyunkwan University)
ClassificationSegmentationDepth EstimationTransformerSupervised Fine-TuningImageText
🎯 What it does: The research proposes a low-bit quantization task vector (Task Vector Quantization, TVQ) and its improved version, Residual Task Vector Quantization (RTVQ), to significantly reduce storage costs during model fusion.
Task-Aware Prompt Gradient Projection for Parameter-Efficient Tuning Federated Class-Incremental Learning
Hualong Ke (Xiamen University), Yanyun Qu (Xiamen University)
Federated LearningTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes the TPPR (Task-aware Prompt Gradient Projection and Replay) method, which uses visual prompts for parameter-efficient fine-tuning in federated incremental learning, avoiding overall model training and large-scale communication.
Task-Decoupled Bezier Surface Constraint for Uneven Low-Light Image Enhancement
Xingxiang Zhou (Inner Mongolia University), Guanglai Gao (Inner Mongolia University)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes a task-decoupled low-light image enhancement method called BSCNet, which utilizes Bezier surface constraints to achieve brightness smoothness control.
Task-Oriented Human Grasp Synthesis via Context- and Task-Aware Diffusers
An-Lun Liu (National Yang Ming Chiao Tung University), Yi-Ting Chen (National Yang Ming Chiao Tung University)
Data SynthesisPose EstimationRobotic IntelligenceTransformerDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a task-oriented human grasp synthesis method, constructing a novel dataset that includes three everyday tasks: placing, stacking, and shelving. It generates a task-aware contact map through a two-stage diffusion model, followed by synthesizing hand poses that meet task objectives while avoiding collisions with the environment.
Task-Specific Zero-shot Quantization-Aware Training for Object Detection
Changhao Li (Georgia Institute of Technology), Jianfei Chen (Tsinghua University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A task-specific zero-shot quantization-aware training framework for object detection is proposed, which includes a two-stage process of generating task-specific calibration sets from pre-trained models and incorporating task-specific distillation in quantization training.
TAViS: Text-bridged Audio-Visual Segmentation with Foundation Models
Ziyang Luo (Northwestern Polytechnical University), Junwei Han (Northwestern Polytechnical University)
SegmentationTransformerPrompt EngineeringVideoTextMultimodalityAudio
🎯 What it does: This paper proposes a multi-modal foundation model-based audio-video segmentation framework called TAViS, which integrates ImageBind and SAM2, achieving alignment and prompting of audio, visual, and text through a text bridge.
TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation
Amin Karimi Monsefi (Ohio State University), Cheng Zhang (Texas A&M University)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper designs and trains a hierarchical taxonomy-driven diffusion model named TaxaDiffusion, aimed at generating high-quality images of fine-grained animal species under few-shot conditions.
TCFG: Truncated Classifier-Free Guidance for Efficient and Scalable Text-to-Image Acceleration
Xiaomeng Fu (Institute of Information Engineering), Jia Li (University of Hong Kong)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: The TCFG method is proposed, which reduces the computational load of Classifier-Free Guidance (CFG) by truncating text guidance after detecting semantic convergence in images, achieving acceleration.
Teaching AI the Anatomy Behind the Scan: Addressing Anatomical Flaws in Medical Image Segmentation with Learnable Prior
Young Seok Jeon (National University of Singapore), Mengling Feng (National University of Singapore)
SegmentationConvolutional Neural NetworkBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a multi-organ segmentation network named AIC-Net, which utilizes learnable anatomical priors and adaptively matches patient anatomy through a differentiable affine + TPS deformation in a global-local cascading structure, thereby guiding the decoder to generate segmentation results that are more anatomically accurate.
Teaching VLMs to Localize Specific Objects from In-context Examples
Sivan Doveh (IBM Research), M. Jehanzeb Mirza (MIT CSAIL)
Object DetectionObject TrackingTransformerSupervised Fine-TuningVision Language ModelVideo
🎯 What it does: A data-driven fine-tuning method called IPLoc is proposed, which constructs context learning dialogues based on video tracking data, enabling visual language models to accurately locate specified objects with a small number of examples.
TeEFusion: Blending Text Embeddings to Distill Classifier-Free Guidance
Minghao Fu (Nanjing University), Kaifu Zhang (Alibaba International Digital Commerce Group)
GenerationComputational EfficiencyKnowledge DistillationDiffusion modelImageText
🎯 What it does: A distillation method called TeEFusion is proposed, which integrates text embedding fusion guided by the teacher model CFG, achieving CFG effects with a single forward inference.
Teeth Reconstruction and Performance Capture Using a Phone Camera
Weixi Zheng (Tsinghua University), Feng Xu (SenseTime Research)
RestorationSegmentationSupervised Fine-TuningImageVideoMesh
🎯 What it does: Using a single mobile phone camera to achieve personalized dental geometry reconstruction and facial motion capture involving teeth.
TeethGenerator: A two-stage framework for paired pre- and post-orthodontic 3D dental data generation
Changsong Lei (Tsinghua University), Yong-Jin Liu (Tsinghua University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderPoint Cloud
🎯 What it does: A two-stage framework called TeethGenerator is proposed, capable of synthesizing paired 3D point clouds of orthodontic anterior and posterior teeth.
Teleportraits: Training-Free People Insertion into Any Scene
Jialu Gao (Carnegie Mellon University), Fernando De La Torre (Carnegie Mellon University)
Image TranslationGenerationDiffusion modelImageBenchmark
🎯 What it does: A training-free Teleportraits pipeline is proposed, which automatically inserts portraits into any background using a single reference image while maintaining high-quality fusion of identity and background.
TemCoCo: Temporally Consistent Multi-modal Video Fusion with Visual-Semantic Collaboration
Meiqi Gong (Wuhan University), Jiayi Ma (Wuhan University)
RestorationData SynthesisKnowledge DistillationConvolutional Neural NetworkOptical FlowVideoMultimodality
🎯 What it does: A multi-modal video fusion framework called TemCoCo is proposed, which takes into account visual quality, semantic accuracy, and temporal consistency.
Temperature in Cosine-based Softmax Loss
Takumi Kobayashi (National Institute of Advanced Industrial Science and Technology)
ClassificationOptimizationContrastive LearningImage
🎯 What it does: This study investigates how to adaptively estimate the temperature parameter of the cos-based softmax loss and proposes a least squares method based on geometric projection.
Temporal Overlapping Prediction: A Self-supervised Pre-training Method for LiDAR Moving Object Segmentation
Ziliang Miao (University of Hong Kong), Fu Zhang (University of Hong Kong)
Object DetectionSegmentationAutonomous DrivingContrastive LearningPoint Cloud
🎯 What it does: A self-supervised pre-training method called TOP is proposed, which learns the spatiotemporal representation of LiDAR moving object segmentation through the prediction of occupancy states at temporally overlapping points.
Temporal Rate Reduction Clustering for Human Motion Segmentation
Xianghan Meng (Beijing University of Posts and Telecommunications), Chun-Guang Li (Beijing University of Posts and Telecommunications)
SegmentationRepresentation LearningVideo
🎯 What it does: This paper proposes a temporal subspace clustering method that jointly learns structured representations and affinity matrices—Temporal Rate Reduction Clustering (TR²C)—for unsupervised segmentation of video frames into different human motion segments.
Temporal Unlearnable Examples: Preventing Personal Video Data from Unauthorized Exploitation by Object Tracking
Qiangqiang Wu (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
Object TrackingTransformerDiffusion modelContrastive LearningVideo
🎯 What it does: This paper proposes a video perturbation technique called Temporal Unlearnable Examples (TUEs) to prevent unauthorized use of personal video data in visual object tracking (VOT) model training.
Temporal-aware Query Routing for Real-time Video Instance Segmentation
Zesen Cheng (Peking University), Jie Chen (Peking University)
Object DetectionSegmentationTransformerVideo
🎯 What it does: A Temporal-aware Query Routing (TAR) mechanism is proposed, which enhances the real-time performance of video instance segmentation by inserting learnable routers before the Transformer decoder layers, allowing for the skipping of unnecessary decoding layers.
Tensor-aggregated LoRA in Federated Fine-tuning
Zhixuan Li (Nanjing University of Science and Technology), Jinhui Tang (Nanjing Forestry University)
Federated LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In a federated learning environment, a parameter-efficient fine-tuning method for large language models is proposed, introducing a Tensor Aggregation LoRA (Te-LoRA) framework.
TeRA: Rethinking Text-guided Realistic 3D Avatar Generation
Yanwen Wang (Nanjing University), Hao Zhu (Nanjing University)
GenerationData SynthesisKnowledge DistillationDiffusion modelGaussian SplattingImageText
🎯 What it does: A text-driven 3D human avatar generation framework called TeRA based on latent diffusion models is proposed, which can quickly, realistically, and editably generate 3D human figures.
TerraMind: Large-Scale Generative Multimodality for Earth Observation
Johannes Jakubik (IBM Research), Nicolas Longépé (European Space Agency)
SegmentationGenerationTransformerDiffusion modelImageMultimodality
🎯 What it does: This paper proposes and implements TerraMind, a generative multimodal Earth observation model that supports arbitrary modal inputs and outputs, and evaluates it on multiple benchmark tasks.
TESPEC: Temporally-Enhanced Self-Supervised Pretraining for Event Cameras
Mohammad Mohammadi (University of Toronto), Igor Gilitschenski (University of Toronto)
Object DetectionSegmentationDepth EstimationRecurrent Neural NetworkTransformerContrastive LearningImageVideo
🎯 What it does: A self-supervised pre-training framework for event cameras, TESPEC, is proposed, which utilizes multi-segment event accumulation to generate pseudo-gray videos as reconstruction targets to mine long-term spatiotemporal information.
Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates
Kecheng Chen, Haoliang Li (City University of Hong Kong)
SegmentationDomain AdaptationComputational EfficiencyTransformerBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A testing adaptation framework is proposed on MedSAM that does not require parameter updates, utilizing latent representation refinement combined with distribution approximated CRF and entropy minimization to achieve self-supervised adaptation.
Test-Time Prompt Tuning for Zero-Shot Depth Completion
Chanhwi Jeong (Gwangju Institute of Science and Technology), Hae-Gon Jeon (Yonsei University)
RestorationDepth EstimationPrompt EngineeringImagePoint Cloud
🎯 What it does: A zero-shot depth completion method based on visual prompts is proposed under the conditions of having only RGB images and sparse depth measurements.
Test-Time Retrieval-Augmented Adaptation for Vision-Language Models
Xinqi Fan (Manchester Metropolitan University), Mubarak Shah (Chinese University of Hong Kong)
RetrievalDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A training-free Test-time Retrieval-Augmented Adaptation (TT-RAA) framework is proposed, which utilizes a Streaming Gaussian Mixture Database (SMGD) to estimate the test distribution in real-time and enhances the inference performance of CLIP through a Retrieval-Augmented Module (MRA).
Text Embedding Knows How to Quantize Text-Guided Diffusion Models
Hongjae Lee (Korea University), Seung-Won Jung (Korea University)
GenerationComputational EfficiencyDiffusion modelImageText
🎯 What it does: A dynamic bit-width quantization method called QLIP is proposed for text-guided diffusion models, which predicts image quality using text prompts and determines the quantization bit-width for each time step and layer accordingly, thereby reducing computational overhead while maintaining image quality.
Text-guided Visual Prompt DINO for Generic Segmentation
Yuchen Guan (Tsinghua University), Chen Li (Tencent Inc.)
SegmentationTransformerPrompt EngineeringImageText
🎯 What it does: Proposes the Prompt-DINO model, which utilizes text and visual prompts for general segmentation, supporting single/multiple prompts and cross-image inference.
Text-IRSTD: Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes
Feng Huang (Fuzhou University), Liqiong Chen (Fuzhou University)
Object DetectionConvolutional Neural NetworkVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a semantic text-guided infrared small target detection framework called Text-IRSTD, which significantly improves target detection and contour recovery in complex scenes.
Text-to-Any-Skeleton Motion Generation Without Retargeting
Qingyuan Liu (University of Chinese Academy of Sciences), Jinbao Wang (Shenzhen University)
GenerationData SynthesisPose EstimationOptimizationTransformerAuto EncoderTextPoint Cloud
🎯 What it does: Developed the OmniSkel framework, achieving text-driven arbitrary skeleton motion generation without redirection.
Text2Outfit: Controllable Outfit Generation with Multimodal Language Models
Yuanhao Zhai (State University of New York at Buffalo), David Doermann (State University of New York at Buffalo)
GenerationRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningTextMultimodality
🎯 What it does: This paper proposes the Text2Outfit framework, which can generate complete outfits based on text prompts, supporting both text-to-outfit and seed-to-outfit recommendation methods.
Text2VDM: Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting
Hengyu Meng (Hong Kong University of Science and Technology), Zeyu Wang (Hong Kong University of Science and Technology)
GenerationPrompt EngineeringDiffusion modelTextMesh
🎯 What it does: This paper proposes a method for generating vector displacement maps (VDM) that can be directly used for 3D asset creation from text, allowing for fine control over surface details and geometric structures.
TextMaster: A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control
Zhenyu Yan (Alibaba Group), Zhu Hangcheng
Image TranslationGenerationDiffusion modelImageText
🎯 What it does: A unified framework named TextMaster is proposed, capable of accurately editing text in different image scenes, ensuring character shape and layout while supporting controllable styles such as font, color, and gradients.
TextSSR: Diffusion-based Data Synthesis for Scene Text Recognition
Xingsong Ye (Fudan University), Zhineng Chen (Fudan University)
RecognitionGenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes TextSSR, a scene text recognition training data synthesis pipeline based on diffusion models, capable of generating high-quality synthetic text images on a large scale while maintaining text accuracy, authenticity, and scalability.
Textured 3D Regenerative Morphing with 3D Diffusion Prior
Songlin Yang (Nanyang Technological University), Xingang Pan (Nanyang Technological University)
GenerationData SynthesisDiffusion modelPoint CloudMesh
🎯 What it does: A texture 3D deformation method based on 3D diffusion priors is proposed, which can achieve smooth and interpretable shape and texture interpolation for different categories of 3D objects without the need for explicit correspondences.
TF-TI2I: Training-Free Text-and-Image-to-Image Generation via Multi-Modal Implicit-Context Learning In Text-to-Image Models
Teng-Fang Hsiao (National Yang Ming Chiao Tung University), Hong-Han Shuai (National Yang Ming Chiao Tung University)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: A multi-modal implicit context learning framework TF-TI2I is proposed, which allows pre-trained text-to-image models (such as SD3, Flux) to jointly generate and edit images using multiple reference images and text prompts without additional training.
The Best of Both Worlds: Integrating Language Models and Diffusion Models for Video Generation
Aoxiong Yin (Zhejiang University), Siliang Tang (Zhejiang University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoText
🎯 What it does: This paper presents LanDiff, a coarse-to-fine video generation framework that combines large language models and diffusion models.
The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation
Ho Kei Cheng (University of Illinois Urbana-Champaign), Alexander Schwing (University of Illinois Urbana-Champaign)
GenerationData SynthesisOptimizationFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: A conditional optimal transport flow matching (C2OT) method is proposed to enhance the performance of flow matching models in conditional generation tasks.
The Devil is in the Spurious Correlations: Boosting Moment Retrieval with Dynamic Learning
Xinyang Zhou (University of Electronic Science and Technology of China), Wen Li (University of Electronic Science and Technology of China)
RetrievalTransformerVideoText
🎯 What it does: This study proposes a dynamic learning method to address the issue of 'spurious correlation' in video temporal retrieval, combining video synthesis and temporal dynamic enhancement to achieve more accurate text-video temporal matching.
The Inter-Intra Modal Measure: A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models
Laura Niss (MIT Lincoln Laboratory), Theodoros Tsiligkaridis (MIT Lincoln Laboratory)
Domain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposes the Inter-Intra Modal Measure (IIMM) to predict the performance improvement and catastrophic forgetting of dual-encoded visual-language models after fine-tuning without requiring fine-tuning itself.
The Scalability of Simplicity: Empirical Analysis of Vision-Language Learning with a Single Transformer
Weixian Lei (Bytedance), Zilong Huang (Bytedance)
RecognitionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A multi-modal large language model SAIL based on a single Transformer architecture is proposed, which can directly encode and decode raw image patches and text tokens in a unified manner, eliminating the need for traditional pre-trained visual encoders and alignment modules.
The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation
Ruoyu Wang (Wuhan University), Yu Wu (Wuhan University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: By constructing a large-scale noise library and retrieving initial noise that matches user objectives, we achieve training-free and interpretable target-driven image generation.
The Source Image is the Best Attention for Infrared and Visible Image Fusion
Song Wang (North University of China), Zhixun Wang (North University of China)
Image TranslationRestorationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a high-quality fusion of infrared and visible images by utilizing the inherent 'attention characteristics' of infrared images through source image cross-attention modules (I-SCA, V-SCA) and a CBSM auxiliary module.
Thermal Polarimetric Multi-view Stereo
Takahiro Kushida (Ritsumeikan University), Kenichiro Tanaka (Ritsumeikan University)
Depth EstimationNeural Radiance FieldImage
🎯 What it does: A method for 3D shape reconstruction using multi-view long-wave infrared (LWIR) polarized images is proposed.
Think Twice: Test-Time Reasoning for Robust CLIP Zero-Shot Classification
Shenyu Lu (Purdue University), Xiaoqian Wang (Purdue University)
ClassificationObject DetectionTransformerVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: This paper proposes an unsupervised test-time reasoning method called TTR, which utilizes CLIP's own semantic representation for object localization in images and performs zero-shot classification only within the localized areas to mitigate the impact of spurious correlations on CLIP's zero-shot classification.
TikZero: Zero-Shot Text-Guided Graphics Program Synthesis
Jonas Belouadi (University of Mannheim), Simone Ponzetto (University of Mannheim)
GenerationData SynthesisKnowledge DistillationTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper proposes Ti k Zero, a zero-shot text-driven TikZ graphic program synthesis method that achieves this by using visual representations as a bridge.
Tile-wise vs. Image-wise: Random-Tile Loss and Training Paradigm for Gaussian Splatting
Xiaoyu Zhang, Guofeng Zhang
OptimizationGaussian SplattingImage
🎯 What it does: Proposes a Random Tile Loss (RTLoss) and a Tile-based training paradigm, optimizing the rendering and geometric quality of 3D Gaussian splatting by combining randomly sampled tiles with multi-view structural information.
Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images
Elena Buglakova (European Molecular Biology Laboratory), Anna Kreshuk (European Molecular Biology Laboratory)
SegmentationBiomedical Data
🎯 What it does: This study investigates the stitching artifacts produced by sliding window inference in large-size biological image segmentation and addresses them by introducing BatchRenorm.
Time-Aware Auto White Balance in Mobile Photography
Mahmoud Afifi (Samsung Electronics), Michael S. Brown (Samsung Electronics)
Convolutional Neural NetworkImage
🎯 What it does: A lightweight automatic white balance estimation method is proposed by integrating metadata such as timestamps, geographic locations, ISO, and shutter speed captured by mobile phones.
TimeBooth: Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging
Zepeng Su (South China University of Technology), C.L.Philip Chen
Image TranslationGenerationTransformerDiffusion modelImage
🎯 What it does: Proposes TimeBooth, which utilizes image-based age prompts to achieve personalized and diverse facial aging generation, and designs a decoupled representation with an age adapter;
TimeExpert: An Expert-Guided Video LLM for Video Temporal Grounding
Zuhao Yang (Nanyang Technological University), Song Bai (ByteDance Inc.)
GenerationRetrievalTransformerLarge Language ModelMixture of ExpertsVision Language ModelVideoText
🎯 What it does: This paper proposes a video temporal localization task-oriented expert-guided large language model called TimeExpert, which utilizes dynamic expert routing to decompose and specialize the handling of subtasks such as timestamps, saliency scores, and text generation.
TimeFormer: Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction
Dadong Jiang (Tianjin University), Tie Qiu (Tianjin University)
RestorationOptimizationTransformerGaussian SplattingVideoPoint Cloud
🎯 What it does: Proposes TimeFormer, a pluggable Transformer module for implicitly learning the temporal relationships of 3D Gaussian points in dynamic scene reconstruction;
Timestep-Aware Diffusion Model for Extreme Image Rescaling
Ce Wang (Wuhan University), Zhenzhong Chen (Wuhan University)
RestorationGenerationSuper ResolutionDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes the Timestep-Aware Diffusion Model (TADM), which performs extreme image resampling in the latent space of a pre-trained VAE and utilizes pre-trained Stable Diffusion for one-time denoising enhancement, achieving a bidirectional mapping from high resolution to low resolution (16×/32×).
TinyViM: Frequency Decoupling for Tiny Hybrid Vision Mamba
Xiaowen Ma (Huawei Noah's Ark Lab), Xinghao Chen (Huawei Noah's Ark Lab)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: A lightweight visual Mamba backbone, TinyViM, is designed and implemented, utilizing a frequency-decoupled Laplace Mixer and a frequency gradient guidance (Inception) mechanism to handle low-frequency global context and high-frequency local details, achieving efficient global modeling and high-frequency retention.
TIP-I2V: A Million-Scale Real Text and Image Prompt Dataset for Image-to-Video Generation
Wenhao Wang (University of Technology Sydney), Yi Yang (Zhejiang University)
GenerationData SynthesisLarge Language ModelPrompt EngineeringDiffusion modelImageVideoTextBenchmark
🎯 What it does: This paper presents the TIP-I2V dataset, which contains over 1.7 million text + image prompts provided by real users, along with corresponding videos generated by five mainstream image-to-video models, aiming to fill the gap in the lack of dedicated prompt datasets in this field.
TITAN-Guide: Taming Inference-Time Alignment for Guided Text-to-Video Diffusion Models
Christian Simon (Sony Group Corporation), Yuki Mitsufuji (Sony Group Corporation)
GenerationData SynthesisOptimizationDiffusion modelVideoMultimodalityAudio
🎯 What it does: We propose TITAN-Guide, a training-free guidance method based on forward gradient for achieving high-quality alignment in memory-constrained text-to-video diffusion models.
TITAN: Query-Token based Domain Adaptive Adversarial Learning
Tajamul Ashraf (Mohamed bin Zayed University of Artificial Intelligence), Janibul Bashir (National Institute of Technology Srinagar)
Object DetectionDomain AdaptationKnowledge DistillationTransformerImageBiomedical Data
🎯 What it does: A TITAN framework is proposed for source-free domain adaptive object detection, which divides target domain samples into easy/difficult subsets based on detection variance, and incorporates query-token adversarial learning in a student-teacher model to achieve domain alignment.
TLB-VFI: Temporal-Aware Latent Brownian Bridge Diffusion for Video Frame Interpolation
Zonglin Lyu (University of Central Florida), Chen Chen (University of Central Florida)
RestorationGenerationData SynthesisDiffusion modelAuto EncoderOptical FlowVideo
🎯 What it does: A Temporal-Aware Latent Brownian Bridge Diffusion (TLB-VFI) model is proposed for high-quality video frame interpolation, balancing temporal information extraction and efficient sampling.
To Label or Not to Label: PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models
Julia Machnio (Pioneer Centre for AI University of Copenhagen), Mostafa Mehdipour Ghazi (Pioneer Centre for AI University of Copenhagen)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes the PALM model, which predicts sample efficiency and learning curves in the active learning process through interpretable mathematical parameters.
ToF-Splatting: Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration
Andrea Conti (University of Bologna), Stefano Mattoccia (University of Bologna)
Depth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A sparse ToF depth SLAM framework named ToF-Splatting is proposed, utilizing 3D Gaussian Splatting (3DGS) for dense map reconstruction, and combining extremely sparse ToF measurements with multi-view geometry and monocular cues through a multi-frame fusion module to generate dense depth maps.
TOGA: Temporally Grounded Open-Ended Video QA with Weak Supervision
Ayush Gupta (SRI International), Susmit Jha (United States Military Academy)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: Proposes the TOGA framework, which completes open-ended answer generation for video question answering while also providing temporal localization.
Token Activation Map to Visually Explain Multimodal LLMs
Yi Li (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper studies and proposes the Token Activation Map (TAM) for visualizing the activation information of multimodal large language models (MLLM) during the step-by-step generation of multi-word outputs.
Token-Efficient VLM: High-Resolution Image Understanding via Dynamic Region Proposal
Yitong Jiang (NVIDIA), Sifei Liu (NVIDIA)
RecognitionObject DetectionSegmentationComputational EfficiencyTransformerVision Language ModelImageMultimodality
🎯 What it does: The TEVA framework proposes a visual language model based on dynamic region sampling, capable of efficiently capturing details and maintaining global context in images of any resolution;
TokensGen: Harnessing Condensed Tokens for Long Video Generation
Wenqi Ouyang (Nanyang Technological University), Xingang Pan (Nanyang Technological University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Proposes the TokensGen framework, which utilizes compressed video tokens to achieve continuous video generation of up to one minute.
TokenUnify: Scaling Up Autoregressive Pretraining for Neuron Segmentation
Yinda Chen (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
SegmentationTransformerImageBiomedical Data
🎯 What it does: By constructing a self-supervised pre-training framework TokenUnify that includes random token prediction, next-token prediction, and next-all-token prediction on EM images, and utilizing the Mamba network for efficient modeling of long sequences, fine-grained neuron segmentation is performed based on this.
ToolVQA: A Dataset for Multi-step Reasoning VQA with External Tools
Shaofeng Yin (Peking University), Yang Liu (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A real-world multimodal tool usage question-answering dataset, ToolVQA, has been constructed, and a ToolEngine generation pipeline for automatically generating multi-step reasoning samples has been proposed.
Top2Pano: Learning to Generate Indoor Panoramas from Top-Down View
Zitong Zhang (University of Louisville), Rui Yu (University of Louisville)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: An end-to-end framework called Top2Pano is proposed to generate high-quality 360° indoor panoramas from 2D top-down views.
TopicGeo: An Efficient Unified Framework for Geolocation
Xin Wang (Xidian University), Shuiping Gou (Xidian University)
RetrievalOptimizationTransformerContrastive LearningImage
🎯 What it does: A unified TopicGeo framework is proposed to achieve a complete geographic localization process that retrieves and matches large-scale reference images directly from query images.
TopoTTA: Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation
Jiale Zhou (Zhejiang University), Yefeng Zheng (Westlake University)
SegmentationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A testing adaptation framework specifically designed for tubular structure segmentation, TopoTTA, can continuously improve model predictions on the target domain without labels, maintaining and enhancing the connectivity and segmentation accuracy of pipeline structures.
TorchAdapt: Towards Light-Agnostic Real-Time Visual Perception
Khurram Azeem Hashmi (German Research Center for Artificial Intelligence), Muhammad Zeshan Afzal
Object DetectionSegmentationDomain AdaptationConvolutional Neural NetworkContrastive LearningImageVideo
🎯 What it does: We propose TorchAdapt, a real-time, illumination-invariant adaptive feature enhancement framework that can improve the performance of various high-level visual tasks under different lighting conditions.
TOTP: Transferable Online Pedestrian Trajectory Prediction with Temporal-Adaptive Mamba Latent Diffusion
Ziyang Ren (Xi'an Jiaotong University), Huan Li (Xi'an Jiaotong University)
Object TrackingGenerationKnowledge DistillationRecurrent Neural NetworkDiffusion modelTime SeriesSequential
🎯 What it does: This paper proposes the Transferable Online Trajectory Prediction task (TOTP) and the Temporal-Adaptive Mamba Latent Diffusion (TAMLD) model based on Mamba, achieving synchronous prediction of all pedestrian trajectories under variable-length observations in real-time scenarios, and supporting transfer between different tasks.
Toward Better Out-painting: Improving the Image Composition with Initialization Policy Model
Xuan Han (Tongji University), Mingyu You (Tongji University)
Image HarmonizationRestorationGenerationDiffusion modelImage
🎯 What it does: This paper proposes an Initialization Strategy Model (IPM) that directly predicts low-frequency signals in the foreground condition highlighting filling task, thereby suppressing the interference of random initial noise on image synthesis and significantly improving the quality of background and foreground integration.
Toward Fair and Accurate Cross-Domain Medical Image Segmentation: A VLM-Driven Active Domain Adaptation Paradigm
Hongqiu Wang (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
SegmentationDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: This paper proposes a Fair Active Domain Adaptation (Fair-ADA) paradigm based on Visual-Language Models (VLM) for cross-domain medical image segmentation tasks, achieving both fairness and efficiency in model transfer with a limited number of labeled samples (approximately 5%). The method utilizes VLM to achieve semantic alignment between images and sensitive attributes (such as race and gender) and designs a FairAP sampling strategy that combines attributes with polysemy to ensure that balanced and diverse samples from various subgroups are included in the labeled set. The model trained on the source domain is then fine-tuned to obtain a fairer and better-performing target domain model.
Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts
Chiao-An Yang (Purdue University), Raymond A. Yeh (Purdue University)
Anomaly DetectionPrompt EngineeringVision Language ModelAuto EncoderImageBiomedical DataBenchmark
🎯 What it does: This study proposes an online anomaly detection model for long-tailed distributions without category labels (LTOAD) and establishes a corresponding benchmark for long-tailed online anomaly detection.
Toward Material-Agnostic System Identification from Videos
Yizhou Zhao (Carnegie Mellon University), Min Xu (Carnegie Mellon University)
OptimizationGaussian SplattingVideo
🎯 What it does: The MASIV (Material-Agnostic System Identification from Videos) framework is proposed, which utilizes multi-view videos to achieve material-agnostic identification of object geometry and dynamics through dynamic Gaussian reconstruction, continuous particle trajectory inference, and a differentiable MPM combined with a neural constitutive model.
Towards a 3D Transfer-based Black-box Attack via Critical Feature Guidance
Shuchao Pang (Nanjing University of Science and Technology), Yongbin Zhou (Nanjing University of Science and Technology)
ClassificationAdversarial AttackPoint Cloud
🎯 What it does: This paper proposes a 3D point cloud black-box transfer attack method based on key feature guidance, named CFG, which can generate adversarial point clouds with a high transfer rate without any information about the target model.
Towards a Unified Copernicus Foundation Model for Earth Vision
Yi Wang (Technical University of Munich), Xiao Xiang Zhu (Technical University of Munich)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningImageMultimodalityBenchmark
🎯 What it does: The Copernicus-Pretrain dataset, Copernicus-FM model, and Copernicus-Bench benchmark are proposed, and the effectiveness of the unified base model for surface and atmospheric multi-tasking is validated.
Towards a Universal 3D Medical Multi-modality Generalization via Learning Personalized Invariant Representation
Zhaorui Tan (Shanghai Academy of Artificial Intelligence for Science), Yuan Cheng (Duke Kunshan University)
Domain AdaptationRepresentation LearningContrastive LearningMultimodalityBiomedical DataMagnetic Resonance ImagingPositron Emission Tomography
🎯 What it does: A method is proposed to enhance the generalization of 3D medical multimodal tasks through learning personalized invariant representations.
Towards a Universal Image Degradation Model via Content-Degradation Disentanglement
Wenbo Yang (University of Waterloo), Zhou Wang (University of Waterloo)
RestorationData SynthesisConvolutional Neural NetworkImage
🎯 What it does: A general image degradation model is proposed, capable of automatically extracting and synthesizing mixed uniform and non-uniform degradations.
Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge
Linshen Liu (Johns Hopkins University), Hao Frank Yang (Johns Hopkins University)
Object DetectionAutonomous DrivingComputational EfficiencyMixture of ExpertsMultimodalityPoint Cloud
🎯 What it does: A marginal 3D object detection system EMC2 based on Mixture of Experts (MoE) is proposed, utilizing multimodal fusion of LiDAR and cameras along with scene-adaptive expert scheduling to enhance autonomous driving perception performance.
Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment
Kejia Zhang (Xiamen University), Zhiming Luo (Xiamen University)
Adversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes Debiased High-Confidence Adversarial Training (DHAT), which enhances the robustness and generalization ability of the model by eliminating background feature bias caused by adversarial training.
Towards Annotation-Free Evaluation: KPAScore for Human Keypoint Detection
Xiaoxiao Wang (University of Chinese Academy of Sciences), Yao Zhu (Zhejiang University)
Object DetectionPose EstimationTransformerSupervised Fine-TuningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: An unsupervised keypoint detection evaluation metric called KPAScore is proposed, which uses a two-stage visual language model-based question answering and visual prompting to determine the existence of keypoints and the accuracy of their localization.
Towards Comprehensive Lecture Slides Understanding: Large-scale Dataset and Effective Method
Enming Zhang (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
TransformerSupervised Fine-TuningTextMultimodality
🎯 What it does: This paper constructs a large-scale multi-task lecture slide dataset, LecSlides-370K, and proposes two tasks: lecture summarization and question answering; it also introduces the SlideParser method, which can predict and utilize text relationships.
Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models
Young Kyun Jang (Google DeepMind), Ser-nam Lim (University of Central Florida)
RetrievalRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a Cross-modal Backward-compatible Training (XBT) method that allows newly launched visual-language pre-trained models (VLP) to share retrieval galleries with old models without recalculating all embeddings.
Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning
Fei Zhou (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Domain AdaptationMeta LearningTransformerContrastive LearningImage
🎯 What it does: An absorptive adaptation learning framework is proposed for extreme cross-domain few-shot learning, guiding the base model to absorb knowledge through an expert model.
Towards Efficient General Feature Prediction in Masked Skeleton Modeling
Shengkai Sun (Hefei University of Technology), Meng Wang (University of Science and Technology of China)
RecognitionPose EstimationRetrievalComputational EfficiencyTransformerContrastive LearningVideo
🎯 What it does: In the self-supervised pre-training of skeletal sequence masking, the low-level coordinate reconstruction objective is replaced with multi-scale high-level semantic feature prediction.
Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars
Yifan Zhan (Shanghai Artificial Intelligence Laboratory), Yinqiang Zheng (The University of Tokyo)
GenerationPose EstimationGaussian SplattingVideoMesh
🎯 What it does: The ToMiE method is proposed, which adaptively grows a tree on the SMPL skeleton to accurately model 3D human figures with handheld objects and loose clothing, and supports explicit animation.
Towards Fine-grained Interactive Segmentation in Images and Videos
Yuan Yao (Alibaba Group), Liefeng Bo (Alibaba Group)
SegmentationTransformerImageVideo
🎯 What it does: Proposes the SAM2Refiner framework, utilizing the SAM2 Backbone to achieve fine-grained interactive segmentation for images and videos.
Towards Foundational Models for Single-Chip Radar
Tianshu Huang, Anthony Rowe
Object DetectionSegmentationAutonomous DrivingTransformerMultimodalityPoint Cloud
🎯 What it does: A general radar model (GRT) based on Transformer is proposed and trained, achieving high-quality 3D occupancy, semantic segmentation, and vehicle motion estimation on 1 million raw mmWave radar data (29 hours).
Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product
Paul Albert (Amazon Machine Learning), Ehsan Abbasnejad (Monash University)
TransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A parameter-efficient fine-tuning method based on the Khatri-Rao product, called KRAdapter, is proposed to address the low-rank limitation of LoRA.