π― What it does: For target detection domain adaptation when source data is unavailable, a simpler self-training strategy is proposed and its effectiveness is evaluated against existing complex methods.
π― What it does: Investigated and repaired the single-value defect in the DINOv2 Vision Transformer, proposing the SINDER method which fixes the defect by fine-tuning singular values and incorporating smooth regularization.
π― What it does: Building upon Diffusion Transformers, we propose Scalable Interpolant Transformers (SiT), incorporating improvements such as interpolators, continuous time training, velocity prediction, and adjustable noise coefficients to create more flexible flow/diffusion generation models, achieving superior FID results on ImageNet 256Γ256 and 512Γ512.
Six-Point Method for Multi-Camera Systems with Reduced Solution Space
Banglei Guan (National University of Defense Technology), Laurent Kneip (ShanghaiTech University)
CodePose EstimationImage
π― What it does: Proposed a minimal solver for solving relative pose in multi-camera systems using six point correspondences (PC), enabling 6DOF pose estimation;
Skeleton Recall Loss for Connectivity Conserving and Resource Efficient Segmentation of Thin Tubular Structures
Yannick Kirchhoff (German Cancer Research Center), Klaus H. Maier-Hein
CodeSegmentationComputational EfficiencyConvolutional Neural NetworkImageBiomedical Data
π― What it does: Propose Skeleton Recall Loss, a loss function that enforces connectivity constraints on thin tubular structures through a precomputed tubed skeleton.
π― What it does: Propose a panoramic graph structure combining multi-person human skeletons and object key points, and design a multi-scale spatiotemporal GCN (MP-GCN) for collective action recognition.
π― What it does: This paper defines and quantifies the completeness and balance of the 'role-filler' relationship in text-image generation, systematically investigates the impact of data distribution skew on model generalization using synthetic and real images (What'sUp benchmark), and proposes two types of statistical metrics for visual and language spaces.
π― What it does: Propose SkyMask, a federated learning framework that detects and defends against Byzantine attacks at the server side by utilizing learnable fine-grained masks.
π― What it does: This paper proposes SLEDGE, a generative model-based and rule-driven traffic driving environment synthesis and simulation framework;
π― What it does: Propose the SLIM method, which enhances model robustness against spuriousness by constructing an attention space, using minimal manual attention annotations, and selecting feature-balanced data.
SMILe: Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection
Anay Majee (University of Texas at Dallas), Rishabh Iyer (University of Texas at Dallas)
CodeObject DetectionImage
π― What it does: Propose a loss framework called SMILe based on submodular mutual information and total submodular information, specifically designed to address class confusion and catastrophic forgetting issues in few-shot object detection.
π― What it does: Proposed SNeRV, which utilizes discrete wavelet transform (DWT) to separate low-frequency and high-frequency components, encodes only the low-frequency components, and reconstructs high-frequency details through MFU/HFR. Further, it extends the time-domain DWT to capture motion, achieving high-quality reconstruction of video implicit representations.
π― What it does: Propose a graph-aware neuron-level pruning method called SNP to compress and accelerate ViT models while maintaining attention scores.
Hossein Jafarinia (Sharif University of Technology), Mohammad Hossein Rohban (Sharif University of Technology)
CodeClassificationComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningBiomedical Data
π― What it does: Proposed a MIL-pooling architecture named Snuffy, which utilizes sparse Transformers for efficient classification of whole slide images (WSI), and employs self-supervised pre-training and Adapter for few-shot fine-tuning.
Shuanghao Bai (Xi'an Jiaotong University), Badong Chen (Xi'an Jiaotong University)
CodeDomain AdaptationRepresentation LearningTransformerPrompt EngineeringVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
π― What it does: Propose a soft prompt generation method based on a generative model (SPG), which first learns domain-level soft prompt labels and then uses a conditional GAN (CGAN) to generate instance-level soft prompts for each sample, thereby improving the performance of vision-language models (e.g., CLIP) in domain generalization tasks.
π― What it does: Propose a unified motion planning framework called GUMP based on generative models, which can learn driving scene dynamics, generate diverse future trajectories, and construct new scenarios under different conditions;
Source Prompt Disentangled Inversion for Boosting Image Editability with Diffusion Models
Ruibin Li (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
CodeGenerationDiffusion modelImage
π― What it does: Design a source prompt decoupling inversion method called SPDInv to enhance the editability of text-driven image editing based on diffusion models.
π― What it does: Propose a Disruption-Compensation framework that reconstructs depth maps using sparse radar and camera data. By disrupting and compensating for the stripe-like scanning patterns (LiDAR Distribution Leakage, LDL) that emerge under sparse LiDAR supervision, the method significantly enhances depth prediction quality.
π― What it does: Propose SpatialFormer, a decoder-only visual Transformer that achieves explicit spatial understanding of images through adaptive spatial tokens and bidirectional cross-attention, which can be directly transferred to multiple tasks such as classification, detection, and segmentation.
π― What it does: Proposed a dataset-agnostic spatially varying degradation model for blind image super-resolution, capable of learning independent degradation kernels for each pixel in the image.
Spherical Linear Interpolation and Text-Anchoring for Zero-shot Composed Image Retrieval
Young Kyun Jang (Meta AI), Ser-Nam Lim (University of Central Florida)
CodeRetrievalSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
π― What it does: Propose a zero-shot compositional image retrieval (ZS-CIR) method based on spherical linear interpolation (Slerp), and introduce a text-anchored fine-tuning (TAT) technique to bridge the modality gap between images and text, achieving efficient retrieval without manual annotation.
π― What it does: Proposed an attention-free Spiking Wavelet Transformer (SWformer) that combines spiking neural networks with wavelet transforms to achieve event-driven learning of high-frequency features.
SQ-LLaVA: Self-Questioning for Large Vision-Language Assistant
Guohao Sun (Rochester Institute of Technology), Zhiqiang Tao (Rochester Institute of Technology)
CodeRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes a self-questioning training method that trains visual-language models to proactively generate high-quality questions related to images after receiving them, thereby enhancing cross-modal alignment and understanding capabilities.
π― What it does: This paper proposes a complete framework named SSL-Cleanse for detecting and removing backdoors in self-supervised learning (SSL) encoders without relying on downstream labels or training sets.
ST-LLM: Large Language Models Are Effective Temporal Learners
Ruyang Liu (Peking University), Ge Li (Peking University)
CodeRecognitionTransformerLarge Language ModelVideoText
π― What it does: Propose the ST-LLM model, which directly inputs all spatial-temporal visual tokens into LLM, achieving efficient video understanding through dynamic masking and global-local input.
π― What it does: This study proposes a framework called STAMP for test-time adaptation in the presence of unknown classes, achieving identification and anomaly detection through reliable class-balanced memory and self-weighted entropy minimization.
Florian Fervers (Fraunhofer Institute of Optronics, System Technologies and Image Exploitation), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
StereoGlue: Joint Feature Matching and Robust Estimation
Daniel Barath (ETH Zurich), Marc Pollefeys (ETH Zurich)
CodePose EstimationImagePoint Cloud
π― What it does: Propose a joint feature matching and robust estimation framework called StereoGlue, which generates candidate matches, estimates models, and guides matching through an iterative process using a single-point minimal solver, thereby achieving consistent one-to-one correspondences and model scores.
π― What it does: Proposes a unified and efficient framework for story visualization and completion called StoryImager, supporting bidirectional generation and addressing both story visualization and story completion tasks.
π― What it does: Propose a pruning method called SLS based on hierarchical feature clustering, which prunes redundant layers caused by frozen parameters in PETL models during cross-domain tasks;
π― What it does: This paper proposes a new framework called BalConpas for Continual Panoptic Segmentation (CPS), aiming to address core challenges in continual learning, including knowledge retention and adaptation to new knowledge, class distribution imbalance, and misleading effects caused by incomplete annotations of replay samples.
π― What it does: Proposes an Attention-like Structural Re-parameterization (ASR) based on Stripe Observation, which uses learnable vectors as inputs to the attention module, making attention values converge to constants after training, thereby achieving structural re-parameterization without additional inference costs.
CodeSegmentationGenerationTransformerDiffusion modelAuto EncoderImageBiomedical Data
π― What it does: Designed and implemented Style-Extracting Diffusion Models (STEDM), which can generate diverse images by leveraging style information from unseen images given content (e.g., semantic layouts), and applied these synthetic images to semi-supervised histopathological segmentation tasks.
π― What it does: Propose a StyleTokenizer method that maps the style of a single reference image into the text embedding space, enabling single-image style control without training in Stable Diffusion;
Subspace Prototype Guidance for Mitigating Class Imbalance in Point Cloud Semantic Segmentation
Jiawei Han (Beijing Institute of Technology), Guangzhi Chen (Beijing Institute of Technology)
CodeSegmentationContrastive LearningPoint Cloud
π― What it does: Propose the Subspace Prototype Guidance (SPG) method, which generates class prototypes through an auxiliary branch in point cloud semantic segmentation and achieves mutual supervision with the main network to alleviate the class imbalance problem.
π― What it does: Propose the SUMix method, which learns mixing proportions and models uncertainty in mixed samples to address the label mismatch problem during the Mixup process.
π― What it does: Proposes the SuperFedNAS method, combining federated learning with super networks to achieve local NAS that requires no additional training after a single training session, enabling rapid fulfillment of diverse device inference objectives.
π― What it does: Proposes SuRF, a neural surface reconstruction framework based on surface centers, which achieves high-fidelity surface reconstruction under sparse multi-view images.
π― What it does: A part-based point cloud completion framework is studied, which leverages complete synthetic point clouds and incomplete real point clouds for domain adaptation, enabling the model to complete real-world point clouds without requiring complete real annotations.
π― What it does: Proposes a 3D point cloud completion method based on Gaussian spherical template-guided coarse-to-fine template generation and correspondence-pooling query generator.
π― What it does: Proposed and implemented T-Rex2, an open-set object detection model that can collaborate through text and visual prompts, supporting multiple prompt methods and achieving zero-shot object detection within a single framework.
π― What it does: By incorporating a local diffusion process into pre-trained diffusion models, first using PatchCore for OOD region detection, then performing parallel generation (branching) separately on OOD and IND regions, and finally fusing predictions (fusion), thus reducing structural hallucinations in image translation.
π― What it does: This paper proposes a zero-shot OOD detection method based on Text Prompt Augmentation (TAG), leveraging CLIP's multimodal features to enhance the separation between ID and OOD samples.
π― What it does: This paper proposes a deformation-driven 3D talking head avatar synthesis framework based on 3D Gaussian scatteringβTalkingGaussianβwhich can generate realistic talking head videos by applying smooth deformations to persistent Gaussian primitives.
Taming CLIP for Fine-grained and Structured Visual Understanding of Museum Exhibits
Ada-Astrid Balauca (INSAIT, Sofia University), Luc Van Gool (ETH Zurich)
CodeRecognitionTransformerVision Language ModelContrastive LearningImageTabular
π― What it does: Studies how to utilize the CLIP model combined with a parsing network to generate structured tabular information from museum exhibit images.
π― What it does: Designed and implemented an image enhancement model called ICELUT, which is entirely based on lookup tables (LUTs), eliminating CNN computations to achieve ultra-low-latency inference.
Task-Driven Uncertainty Quantification in Inverse Problems via Conformal Prediction
Jeffrey Wen (Ohio State University), Phillip Schniter
CodeClassificationRestorationFlow-based ModelBiomedical Data
π― What it does: This paper proposes a task-driven uncertainty quantification framework based on conformal prediction, for evaluating uncertainty in downstream tasks (e.g., soft output classification) from images recovered with limited measurements.
π― What it does: Achieve pose and appearance feature alignment using a frozen pre-trained ControlNet and LoRA, and propose the TCAN framework to realize temporally consistent animation under driven video poses by introducing Temporal ControlNet and Pose-Driven Temperature Map.
Teach CLIP to Develop a Number Sense for Ordinal Regression
Yao DU, Xiaomeng Li (Hong Kong University of Science and Technology)
CodeVision Language ModelContrastive LearningMultimodalityBenchmark
π― What it does: Propose NumCLIP, which leverages CLIP's cross-modal knowledge to learn numerical perception and improve ordinal regression performance
π― What it does: This paper proposes the Teddy framework, which decouples bi-level optimization through Taylor approximation and pre-caches weak teacher models to achieve efficient training for large-scale dataset distillation.
π― What it does: Propose an unsupervised video denoising framework called TAP that leverages a pre-trained image denoiser to remove noise from videos without requiring noisy-clean video pairs;
Temporal Event Stereo via Joint Learning with Stereoscopic Flow
Hoonhee Cho (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
CodeDepth EstimationAutonomous DrivingConvolutional Neural NetworkContrastive LearningOptical FlowTime Series
π― What it does: This paper proposes a temporal event stereo matching framework that significantly improves the stereo matching accuracy of event cameras by jointly learning stereo flow and stereo matching networks, using features and cost volumes from previous time steps for Warping and fusion.
Yuhan Bao (Zhejiang University), Kaiwei Wang (Zhejiang University)
CodeRestorationSuper ResolutionTransformerImageTime Series
π― What it does: Proposed a time-mapping photography method based on an event camera (EvTemMap), which gradually increases light transmittance using a variable aperture in static scenes, records the timestamp of the first positive event (IPE) for each pixel, and converts sparse event streams into dense grayscale images.
π― What it does: Proposed a stereo matching method called TC-Stereo based on temporal consistency, which utilizes semi-dense disparity projection from previous frames for completion, state fusion, and bidirectional iterative refinement in dual spaces (disparity and disparity gradient).
π― What it does: Proposed Text2LiDAR, a text-controlled LiDAR point cloud generation framework capable of converting natural language descriptions into high-quality, controllable 360Β° LiDAR data.
Textual Knowledge Matters: Cross-Modality Co-Teaching for Generalized Visual Class Discovery
Haiyang Zheng (University of Trento), Zhun Zhong (University of Trento)
CodeClassificationRecognitionLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
π― What it does: Propose a two-stage framework called TextGCD, which first constructs image captions through retrieval-based text generation, and then improves the accuracy of general category discovery by utilizing cross-modal co-teaching and alignment.
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
π― What it does: Proposed a logic-rich text-to-image generation task and constructed the TV-Logic dataset along with the baseline UnR-GAN model.
TF-FAS: Twofold-Element Fine-Grained Semantic Guidance for Generalizable Face Anti-Spoofing
Xudong Wang (Xiamen University), Rongrong Ji (Xiamen University)
CodeAnomaly DetectionTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Propose a multimodal framework named TF-FAS for generalized face anti-spoofing, leveraging two fine-grained semantic guidance mechanisms (content and category) to enhance the model's cross-domain generalization capability.
The All-Seeing Project V2: Towards General Relation Comprehension of the Open World
Weiyun Wang (Fudan University), Jifeng Dai (OpenGVLab, Shanghai AI Laboratory)
CodeRecognitionObject DetectionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper proposes the Relation Conversation (ReC) task, constructs a high-quality dataset AS-V2, and designs an evaluation benchmark CRPE. Based on these resources, the All-Seeing Model v2 (ASMv2) was trained, achieving significant improvements in image relation understanding, scene graph generation, and general vision-language tasks.
π― What it does: Propose a semi-supervised domain generalization framework for medical image segmentation, which utilizes statistical individual branches (SIBs) to obtain reliable pseudo-labels, learns domain-invariant features through statistical aggregation branches (SAB), and simulates unknown domains by introducing multi-level perturbations at both image and feature levels.
The First to Know: How Token Distributions Reveal Hidden Knowledge in Large Vision-Language Models?
Qinyu Zhao (Australian National University), Stephen Gould (Australian National University)
CodeSafty and PrivacyTransformerSupervised Fine-TuningVision Language ModelMultimodality
π― What it does: By performing linear probing on the logit distribution of the first token from large vision-language models (LVLMs), the method determines whether the model should answer a question, thereby identifying unanswerable visual questions, jailbreak attacks, and deceptive questions. During the generation process, a decoding strategy based on the probing results is adopted to enhance the safety and reliability of the generated content.
The Hard Positive Truth about Vision-Language Compositionality
Amita Kamath (University of Washington), Ranjay Krishna (University of Washington)
CodeRetrievalRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper constructs an evaluation dataset and training set containing hard positive samples to investigate the compositional reasoning ability of vision-language models for image description.
π― What it does: This paper proposes the NeuSky method, which utilizes sky pixel constraints and an outward-inward visibility network to achieve inverse rendering for outdoor scenes, decoupling geometry, albedo, distant lighting, and sky visibility.
Think before Placement: Common Sense Enhanced Transformer for Object Placement
Yaxuan Qin (Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences)
CodeGenerationTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityChain-of-Thought
π― What it does: Propose a 'think before placing' framework named CSENet, which generates location descriptions using a large multimodal model, then predicts the scale and coordinates of foreground objects to achieve object placement that is more semantically and visually consistent.
This Probably Looks Exactly Like That: An Invertible Prototypical Network
Zachariah Carmichael (University of Notre Dame), Walter Scheirer
CodeClassificationGenerationFlow-based ModelImage
π― What it does: Propose ProtoFlow, a reversible prototype network that combines conceptual neural networks with flow models to achieve joint generation and prediction, where prototypes represent latent space distributions and can be directly mapped back to the data space for visualization.
π― What it does: Studied the upper bound of the spectral norm of the Jacobian in convolutional layers, proposed a new upper bound, and proved that this upper bound can be efficiently and differentiably computed during training.
Tiny Models are the Computational Saver for Large Models
Qingyuan Wang (University College Dublin), Deepu John (University College Dublin)
CodeComputational EfficiencyKnowledge DistillationMixture of ExpertsImage
π― What it does: Propose TinySaver, a framework that dynamically compresses large models by using a pre-trained mini model for early exit before inference;
TIP: Tabular-Image Pre-training for Multimodal Classification with Incomplete Data
Siyi Du (Imperial College London), Chen Qin (Imperial College London)
CodeClassificationData-Centric LearningTransformerAuto EncoderContrastive LearningImageMultimodalityTabularBiomedical Data
π― What it does: Proposed the TIP framework, achieving multi-modal pre-training and downstream classification on tabular and image data with missing values;
π― What it does: Proposed a method called UnlearnDiffAtk for generating adversarial prompts based on the internal 'free' classifier of diffusion models, aimed at evaluating the robustness of safety-driven unlearned diffusion models (unlearned DMs).
TOD3Cap: Towards 3D Dense Captioning in Outdoor Scenes
Bu Jin (Chinese Academy of Sciences), Hao Zhao (Li Auto)
CodeObject DetectionAutonomous DrivingTransformerLarge Language ModelVision Language ModelImageTextMultimodalityPoint Cloud
π― What it does: Proposed the 3D dense description task for outdoor scenes and designed an end-to-end TOD Cap 3 network to output 3D object boxes and natural language descriptions under LiDAR point cloud and panoramic RGB image inputs.
CodeRecognitionSegmentationGenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: A unified, promptable vision foundation model named TAP was constructed, capable of simultaneously performing segmentation, identification, and generating descriptions for any region.
Toward Open Vocabulary Aerial Object Detection with CLIP-Activated Student-Teacher Learning
Yan Li, Wenxian Yu (Tongji University)
CodeObject DetectionKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImage
π― What it does: Propose CastDet, a student-teacher self-learning framework based on the pre-trained RemoteCLIP model, aiming to achieve open-vocabulary object detection (OVD) from a drone perspective, capable of identifying target categories outside the training set without additional annotated data.
π― What it does: This paper proposes three frequency regularization methods (bandwidth-limited input, controllable upsampler, and learnable Lipschitz regularization), which enhance the performance of untrained networks (e.g., DIP) in medical image reconstruction by directly adjusting the network's frequency bias, eliminating dependency on specific architectures.
Towards High-Quality 3D Motion Transfer with Realistic Apparel Animation
Rong Wang (Australian National University), HONGDONG LI
CodeGenerationGraph Neural NetworkMesh
π― What it does: Propose a data-driven 3D motion transfer method that can animate target stylized characters while generating realistic clothing animations.
π― What it does: This paper proposes and studies the Ambient Lighting Normalization (ALN) task, aiming to recover image details in complex multi-light source and self-shadowed scenes.
π― What it does: Propose and systematically evaluate a framework for Latent Masked Image Modeling (Latent MIM) in the latent space, leveraging reconstruction of latent features to learn unsupervised visual representations.
π― What it does: Propose a model-agnostic dataset compression method called HMDC, which generates general synthetic images by using heterogeneous models collaboratively.
Towards Multi-modal Transformers in Federated Learning
Guangyu Sun (University of Central Florida), Chen Chen (University of Central Florida)
CodeFederated LearningRepresentation LearningTransformerMixture of ExpertsImageTextMultimodalityBiomedical Data
π― What it does: Propose the FedCola framework, addressing cross-modal and modality gap issues between single-modal and multi-modal clients in federated learning through a multi-modal Transformer.
π― What it does: Proposed the MOOSA method for Multimodal Open-Set Domain Generalization and Adaptation (MMOSDG/MM-OSDA), which learns cross-modal representations through self-supervised tasks and achieves unknown class detection.
Minkyu Choi (University of Texas at Austin), Sandeep Chinchali (University of Texas at Austin)
CodeRetrievalExplainability and InterpretabilityTransformerVision Language ModelVideoText
π― What it does: Propose a neuro-symbolic video retrieval framework that integrates visual language models with temporal logic and probabilistic automata to real-time locate complex event scenarios in long videos.
π― What it does: This paper proposes an open-world object detection framework based on self-supervised virtual anomaly synthesis (SSOS) to achieve object-level anomaly detection without class labels.
π― What it does: This paper proposes an end-to-end framework that utilizes an event camera to simultaneously enhance low-light videos and remove motion blur.
Towards Reliable Advertising Image Generation Using Human Feedback
Zhenbang Du (Huazhong University of Science and Technology), Jingping Shao (JD)
CodeGenerationData SynthesisConvolutional Neural NetworkTransformerReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Proposed a reliable system for generating advertising images, utilizing a multi-modal RFNet to simulate human review and combining cyclic generation with RFFT-refined diffusion models, significantly improving the proportion of usable images.
π― What it does: This paper proposes a reliable robustness evaluation and fast training method for semantic segmentation models, including a novel attack loss function, attack ensemble, and adversarial training using a robust ImageNet backbone.
π― What it does: Through unpaired event-to-event translation, daytime event data is converted into nighttime events, helping networks that perform well on daytime data maintain performance at night.
π― What it does: This paper proposes a full low-bit quantization method for super-resolution networks. The core idea is to first map images to the differential operator (edge, texture) domain, enhance the image using a quantized CNN in this domain, and then revert the result back to the original domain through a regularized partial differential equation (PDE) solver, achieving high-quality super-resolution under full 4-bit quantization.
Jiabao Wang (Nankai University), Qibin Hou (Nankai University)
CodeObject DetectionAutonomous DrivingPoint Cloud
π― What it does: This paper proposes a Stability Index (SI) metric to measure the temporal stability of 3D object detectors, and develops a Prediction Consistency Learning (PCL) training strategy based on this, significantly improving the consistency of confidence, position, size, and orientation of detectors across consecutive frames.
Irit Chelly (Ben-Gurion University of the Negev), Oren Freifeld (Ben-Gurion University of the Negev)
CodeClassificationSegmentationGenerationImage
π― What it does: This paper proposes a trainable, high-expressive activation function called DiTAC, which achieves significant performance improvements across various tasks.
Training A Small Emotional Vision Language Model for Visual Art Comprehension
Jing Zhang (Hefei University of Technology), Dan Guo (Hefei University of Technology)
CodeClassificationGenerationTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Trained and evaluated a small-scale emotion vision-language model SEVLM for emotion classification and emotion explanation generation in visual artworks.
π― What it does: Propose an untrained composite scene generation method that utilizes layout information to guide Stable Diffusion for multi-object image synthesis.
Training-free Video Temporal Grounding using Large-scale Pre-trained Models
Minghang Zheng (Peking University), Yang Liu (Peking University)
CodeRetrievalTransformerLarge Language ModelVision Language ModelVideoTextChain-of-Thought
π― What it does: Propose a training-free video temporal localization method, which first splits queries into sub-events and infers their order and relationships using a large language model, then performs dynamic and static matching for each sub-event with a vision-language model, ultimately obtaining video segments.
π― What it does: Proposed a single-stage surface anomaly detection method called TransFusion based on transparency diffusion, which simultaneously reconstructs normal appearance and localizes anomalies in one iteration process;
TTD: Text-Tag Self-Distillation Enhancing Image-Text Alignment in CLIP to Alleviate Single Tag Bias
Sanghyun Jo (OGQ), Kyungsu Kim (Massachusetts General Hospital and Harvard Medical School)
CodeKnowledge DistillationTransformerVision Language ModelContrastive LearningImageText
π― What it does: Proposes the text-label self-distillation method TTD, which mitigates the single-label bias in the CLIP model and achieves more fair image-text alignment by fine-tuning with only image-text pairs.