🎯 What it does: This paper studies the cross-domain few-shot classification problem and proposes a lightweight and parameter-efficient feature space adaptation method.
🎯 What it does: A decoupled pre-training framework DP-HOI is proposed, which separately pre-trains the detection and interaction decoders of DETR using object detection and action recognition data, ultimately aimed at improving HOI detection performance.
🎯 What it does: A controllable visual enhancer DDBF is proposed, utilizing cross-modal conditional adversarial learning to achieve brightness enhancement of low-light visible images and infrared-visible fusion.
🎯 What it does: A distribution-aware knowledge prototyping (DKP) framework is proposed, combining instance-level distribution modeling with distribution-guided prototype generation to achieve zero-shot memory lifelong face re-identification.
🎯 What it does: A two-stage label-free fair facial attribute classification method based on generative models is proposed. First, potential biased attributes are detected through a generative model, and then random amplitude generative augmentation is applied to the potential biased attributes of each image, training a representation network that is insensitive to these augmentations to achieve fair predictions.
🎯 What it does: DistriFusion has been developed, an untrained multi-GPU parallel inference algorithm designed to accelerate the generation of high-resolution diffusion models while maintaining image quality.
🎯 What it does: Proposes the DiverGen method, which generates data using multi-level categories, prompts, and model diversity, and conducts instance segmentation training on LVIS;
🎯 What it does: The DMR framework is proposed, utilizing multimodal inputs from RGB frames and event cameras, explicitly decomposed into task-relevant common features and modality-specific noise, followed by using common features to drive RL decision-making.
🎯 What it does: A dense video captioning framework for cross-modal memory retrieval, CM2, is proposed, which enhances event localization and description using external text memory.
DocRes: A Generalist Model Toward Unifying Document Image Restoration Tasks
Jiaxin Zhang (South China University of Technology), Lianwen Jin (South China University of Technology)
CodeRestorationTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes DocRes, a unified model for five document image restoration tasks (dewarping, shadow removal, appearance enhancement, deblurring, and binarization).
🎯 What it does: This paper proposes a Domain-Separated Graph Neural Network (DSGNN) for simultaneously ranking and segmenting multiple targets in images based on saliency.
🎯 What it does: This paper proposes a Domain-Agnostic Mutual Prompting (DAMP) method utilizing CLIP for unsupervised domain adaptation, aligning visual and textual prompts to enhance the transfer of source domain knowledge to the target domain.
Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation
Yeonguk Yu (Gwangju Institute of Science and Technology), Kyoobin Lee (Gwangju Institute of Science and Technology)
CodeDomain AdaptationImage
🎯 What it does: The DPLOT framework is proposed, which first updates network blocks that only affect domain-specific features through domain-specific block selection before training, and then generates high-quality pseudo-labels during online testing using paired views with only horizontal flipping, performing entropy minimization and symmetric cross-entropy training.
Drag Your Noise: Interactive Point-based Editing via Diffusion Semantic Propagation
Haofeng Liu (Singapore Management University), Shengfeng He (Singapore Management University)
CodeDiffusion modelImageBenchmark
🎯 What it does: This paper proposes DragNoise, a point-dragging interactive image editing method based on diffusion models, which utilizes the bottleneck features of U-Net for single-step semantic optimization and directly replaces the bottleneck features in subsequent time steps for efficient propagation.
🎯 What it does: This paper presents DragDiffusion, an interactive point-based image editing framework based on diffusion models, which supports precise drag-and-drop editing on both real images and diffusion-generated images.
🎯 What it does: We propose Drive-WM, a world model capable of generating high-quality, controllable consistent videos from multiple perspectives, and apply it to end-to-end autonomous driving planning.
🎯 What it does: An end-to-end dense relationship Transformer network called DSGG is designed, which utilizes graph-aware queries to predict a complete scene graph in one go, including object categories, bounding boxes, pixel segmentation, and all relationships between objects.
🎯 What it does: A dual-space network is proposed to simultaneously learn category embeddings and object identity embeddings for pose-invariant object recognition and retrieval.
🎯 What it does: Proposes the Dual Prior Unfolding (DPU) model for efficient reconstruction of three-dimensional hyperspectral images from spectral snapshot compressed imaging (SCI).
🎯 What it does: This paper proposes a Dual Prototype Attention (DPA) framework for unsupervised video object segmentation, primarily achieved through two modules: Inter-Modality Attention (IMA) and Inter-Frame Attention (IFA), which facilitate the dense fusion of RGB and optical flow information and the effective propagation of global video context.
Dual-Scale Transformer for Large-Scale Single-Pixel Imaging
Gang Qu (Westlake University), Xin Yuan (Westlake University)
CodeRestorationTransformerImage
🎯 What it does: This paper proposes a deep unfolding network HATNet for large-scale single-pixel imaging (SPI) reconstruction, utilizing the Kronecker SPI model to achieve direct sampling and reconstruction of the entire image.
🎯 What it does: DUSt3R is proposed, an end-to-end model that can directly generate dense 3D reconstruction and camera parameters from uncalibrated, unposed multi-view images.
🎯 What it does: Proposes a method for dynamic adapters and internal prompt tuning to achieve efficient transfer learning of parameters for pre-trained point cloud models.
CodePose EstimationRecurrent Neural NetworkTime Series
🎯 What it does: This paper proposes DynaIP, a real-time full-body pose estimation framework based on six sparse IMUs, which achieves pseudo-velocity regression and pose regression through a two-stage network, and employs part-based learning and global feature fusion.
Dynamic Prompt Optimizing for Text-to-Image Generation
Wenyi Mo (Renmin University of China), Qing Yang (Du Xiaoman Technology)
CodeGenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImageText
🎯 What it does: This paper proposes an automated prompt editing method (Prompt Auto-Editing, PAE), which generates dynamically fine-grained control prompts (DF-Prompt) with weights and temporal ranges through pre-trained language models and reinforcement learning, aiming to enhance the visual aesthetics and semantic consistency of text-to-image generation.
🎯 What it does: The DYSON framework is proposed for online task-free boundary category incremental learning, which first calculates the optimal geometric structure and aligns the feature space to achieve buffer-free incremental learning.
🎯 What it does: A medical image segmentation method is proposed that achieves continuous testing adaptation by learning low-frequency visual prompts for each test image without updating model parameters.
🎯 What it does: This paper redefines the task of locating Earth images taken by astronauts in space capsules as a large-scale image retrieval problem, retrieving the most similar satellite images from a database and using their geographic tags for localization.
🎯 What it does: We propose EasyDrag, a point-controlled interactive image editing framework based on a pre-trained diffusion model, capable of accurately dragging key points in images without the need for LoRA fine-tuning or hand-drawn masks.
🎯 What it does: This paper proposes a panoptic segmentation method for continuous generalization called ECLIPSE, which utilizes visual prompt tuning to fine-tune a frozen base model, achieving continual learning without distillation by only fine-tuning a small number of prompts.
CodeAutonomous DrivingLarge Language ModelNeural Radiance FieldImage
🎯 What it does: The ChatSim system is proposed, which can edit and render editable high-fidelity 3D driving scenes through natural language commands and supports the import of external digital assets.
🎯 What it does: Utilize diffusion models to distill large-scale datasets, generating smaller, representative datasets rich in original data information.
🎯 What it does: This paper proposes Deformable Convolution v4 (DCNv4) as an efficient and deformable sparse convolution operation, replacing the previous DCNv3, forming faster and stronger visual backbone networks such as FlashInternImage.
Efficient Detection of Long Consistent Cycles and its Application to Distributed Synchronization
Shaohan Li (University of Minnesota), Gilad Lerman (University of Minnesota)
CodeOptimizationComputational EfficiencyGraph
🎯 What it does: This paper proposes an efficient long-period consistency detection method called LongSync, which robustly solves the group synchronization problem in global structured light sequencing.
Efficient Hyperparameter Optimization with Adaptive Fidelity Identification
Jiantong Jiang (University of Western Australia), Ajmal Mian (University of Western Australia)
CodeOptimizationHyperparameter SearchTabular
🎯 What it does: FastBO is proposed, an adaptive multi-fidelity Bayesian optimization method that determines suitable low-order precision for each configuration to build surrogate models.
Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring
Xin Gao (China University of Mining and Technology Beijing), Huaping Liu (University of Science and Technology of China)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: A multi-scale network MLWNet based on Single Input Multiple Output (SIMO) is proposed, introducing a learnable discrete wavelet transform (LWN) module for blind motion deblurring.
Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization
Khiem Le (University of Notre Dame), Kok-Seng Wong (VinUniversity)
CodeDomain AdaptationFederated LearningConvolutional Neural NetworkImageBiomedical Data
🎯 What it does: In federated domain generalization, a personalized normalization method gPerXAN is proposed by explicitly assembling Instance Normalization and Batch Normalization, with the addition of regularization guidance, while maintaining data privacy throughout.
🎯 What it does: This study constructs a high-quality facial image dataset named EFHQ for extreme poses and provides multi-task subsets and cross-pose validation benchmarks based on it.
EgoThink: Evaluating First-Person Perspective Thinking Capability of Vision-Language Models
Sijie Cheng (Tsinghua University), Yang Liu (Tsinghua University)
CodeLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Construct the EgoThink benchmark to evaluate the multi-dimensional thinking abilities of Vision-Language models in terms of objects, activities, localization, reasoning, prediction, and planning from a first-person perspective.
EGTR: Extracting Graph from Transformer for Scene Graph Generation
Jinbae Im (NAVER Cloud AI), Seunghyun Park (NAVER)
CodeObject DetectionGenerationTransformerGraph
🎯 What it does: This paper proposes a lightweight one-stage scene graph generation model, EGTR, which directly constructs an inter-object relationship graph using the weights and query/key vectors from the multi-head self-attention in the DETR decoder, and predicts predicates with a shallow relationship classification head; it also introduces adaptive smoothing and connectivity prediction as auxiliary training.
Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld
Yijun Yang (Southern University of Science and Technology), Yuhui Shi (Southern University of Science and Technology)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A multimodal embodied agent named EMMA has been constructed, utilizing LLM to train VLM in executing embodied tasks in the visual world based on expert behavior in a parallel text world.
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Developed the EmoVIT framework, which utilizes emotion visual instruction data generated by GPT-4 for instruction tuning in visual emotion recognition;
🎯 What it does: This paper proposes a 'Model-Aware Resampling' method (LMAR), which predicts compensation convolution kernels specific to each UHD image and downsampling scale, embedding the compensation information into the downsampled image to achieve a synergistic optimization of resampling and enhancement models.
🎯 What it does: This paper proposes a cross-class image mixing (Diff-Mix) data augmentation method based on diffusion models, utilizing the fine-tuned Stable Diffusion to interpolate images from different categories, thereby enhancing background diversity while maintaining the authenticity of foreground objects, to improve domain-specific image classification tasks.
🎯 What it does: This paper proposes a sample-level multimodal contribution evaluation metric based on Shapley values, and implements a gain resampling strategy for low-contribution modalities based on this evaluation, significantly enhancing the collaborative effect of multimodal learning models.
🎯 What it does: This paper proposes a technique to alleviate the bias towards the compressed domain when enhancing the quality of compressed images; by conditioning the discriminator of the generative adversarial network and introducing domain divergence regularization, the enhanced images are made closer to the original domain, thereby improving perceptual quality and the authenticity of the compressed images.
🎯 What it does: Proposes the SMER (Stochastic Mini-batch Ensemble Reweighting with Reinforcement Learning) method, which gradually generates transferable adversarial examples by leveraging the diversity of multiple models.
🎯 What it does: By applying random equivariant transformations (rotation, flipping, translation) to the trained denoiser during the inference phase and performing inverse transformations on the output, an equivariant denoiser was achieved, significantly improving reconstruction stability and image quality under implicit prior frameworks such as PnP, RED, and ULA.
🎯 What it does: This paper proposes a self-supervised learning framework ES3, which gradually learns shared, unique, and collaborative information between audio and video using an 'evolution' strategy to obtain robust multimodal speech representations.
ESCAPE: Encoding Super-keypoints for Category-Agnostic Pose Estimation
Khoi Duc Nguyen (University of Wisconsin Madison), Gim Hee Lee (National University of Singapore)
CodePose EstimationImage
🎯 What it does: Proposes the ESCAPE framework, which utilizes super keypoint priors to achieve category-agnostic pose estimation, avoiding the use of keypoint identifiers.
🎯 What it does: An end-to-end network based on Transformer is proposed to estimate extreme 3D rotations (only estimating pitch and yaw, assuming roll is 0) when the overlap between image pairs is minimal or nonexistent.
Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning
Rui Zhao (Xi'an Jiaotong University), Bo Dong (Xi'an Jiaotong University)
CodeClassificationSupervised Fine-TuningImage
🎯 What it does: This paper proposes to enhance the posterior estimation of noisy classes by incorporating instance-based clipping multi-label part supervision in noisy label learning.
🎯 What it does: A visual target tracking framework based on hierarchical knowledge distillation for event cameras is designed, achieving high-speed and low-latency tracking using only event signals, and a high-resolution event tracking dataset, EventVOT, is proposed.
Ethan Elms (University of Adelaide), Tat-Jun Chin (Stanford University)
CodeObject TrackingPose EstimationDepth EstimationOptimizationSimultaneous Localization and MappingPoint CloudBenchmark
🎯 What it does: This paper proposes an event camera-based structure-from-orbit (eSfO) method for reconstructing the sparse 3D structure of rotating objects and estimating rotational trajectory parameters when observed by a static event camera.
🎯 What it does: This paper studies a few-shot image generation framework called F2DGAN that achieves precise fusion through feature distribution matching.
ExMap: Leveraging Explainability Heatmaps for Unsupervised Group Robustness to Spurious Correlations
Rwiddhi Chakraborty (UiT The Arctic University of Norway), Michael C. Kampffmeyer (UiT The Arctic University of Norway)
CodeClassificationExplainability and InterpretabilityImage
🎯 What it does: This paper proposes ExMap, a two-stage unsupervised method that first uses Layer-wise Relevance Propagation (LRP) to obtain interpretable heatmaps, then clusters to generate pseudo group labels, which are subsequently used in existing group robust learning strategies to enhance the model's robustness against spurious features.
🎯 What it does: This paper proposes a dataset distillation method based on distribution matching, introducing class centering constraints and covariance matching constraints to enhance the class distinguishability and distribution matching accuracy of synthetic data.
Exploring Regional Clues in CLIP for Zero-Shot Semantic Segmentation
Yi Zhang (Beihang University), Shi-Min Hu (Tsinghua University)
CodeSegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a CLIP-based single-stage zero-shot semantic segmentation framework called CLIP-RC, which utilizes region-level bridging and a recovery decoder to achieve the transfer from image-level knowledge to pixel-level semantics.
Exploring the Potential of Large Foundation Models for Open-Vocabulary HOI Detection
Ting Lei (Peking University), Yang Liu (Peking University)
CodeRecognitionObject DetectionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A human-computer interaction detection framework CMD-SE is proposed for open vocabulary, utilizing multi-layer visual feature conditional matching and fine-grained semantic enhancement of body part descriptions generated by LLM to achieve interaction recognition.
🎯 What it does: A completely unsupervised long-range point cloud registration method called EYOC is proposed, which utilizes continuous LiDAR sequences to adaptively learn features and achieve registration through self-labeling.
🎯 What it does: This paper proposes a weakly supervised instance segmentation method using extreme points (the topmost, leftmost, bottommost, and rightmost points of the target). First, it trains a pseudo-label generator using extreme points, and then uses the generated high-quality pseudo-masks to train a conventional instance segmentation network.
🎯 What it does: The study addresses the fairness issue in federated learning scenarios with domain skew, proposing the FedHEAL framework to alleviate performance unfairness by filtering out unimportant parameters and adjusting aggregation weights.
🎯 What it does: This paper proposes a synthetic image detector that uses noise images and reconstructed images obtained from the DDIM inversion of Stable Diffusion, along with the original images as input.
🎯 What it does: This paper proposes the Dual-SAM framework, designed specifically for marine animal segmentation tasks, using a dual-branch SAM encoder, cross-layer adapters, multi-layer coupled prompts, dilated fusion attention, and cross-connected prediction.
🎯 What it does: A single-step ODE solver (AMED-Solver) and its plugin are proposed, utilizing the geometric properties of sampling trajectories that almost lie in a two-dimensional subspace to eliminate truncation errors by learning the mean direction, thus achieving high-quality image generation with only about 5 function evaluations (NFE).
🎯 What it does: Using graph signal processing and random spectral sampling on 3D correspondence graphs, the FastMAC algorithm is proposed, significantly accelerating the registration of the maximum clique (MAC) to real-time levels while maintaining a high registration success rate.
🎯 What it does: To address the issues of noise and keypoint errors in sparse feature matching, we propose FC-GNN, a graph neural network capable of jointly filtering out anomalous matches and calibrating the precision of the remaining matches.
🎯 What it does: The Feature Calibration and Separation (FCS) method is proposed for No-Example Class Incremental Learning (NECIL), which alleviates catastrophic forgetting by calibrating old class prototypes and enhancing feature separation between old and new classes.
🎯 What it does: This paper proposes an online instance feature re-embedding framework that utilizes the Re-embedded Regional Transformer (R2 Transformer) to re-embed offline extracted instance features in the multi-instance learning (MIL) process, significantly improving the performance of computational pathology tasks.
FedAS: Bridging Inconsistency in Personalized Federated Learning
Xiyuan Yang (Wuhan University), Mang Ye (Wuhan University)
CodeClassificationFederated LearningImage
🎯 What it does: Proposes the FedAS framework to address the intra-client and inter-client inconsistency issues in personalized federated learning, enhancing the collaborative training of localized shared parameters and personalized heads.
🎯 What it does: Proposed the Fed-GCD task and designed the AGCL framework, combining federated learning with learnable GMM to achieve universal category discovery without shared data.
🎯 What it does: A new scenario of Heterogeneous Client Federated Multi-Task Learning (HC-FMTL) is proposed, and the FedHCA2 framework is designed to achieve personalized model federated training for clients under different task settings.
FedSelect: Personalized Federated Learning with Customized Selection of Parameters for Fine-Tuning
Rishub Tamirisa (University of Illinois Urbana-Champaign), Aviv Shamsian (Bar-Ilan University)
CodeFederated LearningImage
🎯 What it does: An algorithm called FedSelect is proposed for adaptive selection of personalized sub-networks in federated learning, which can gradually expand the personalized parameters of each client while maintaining global sharing.
CodeAutonomous DrivingKnowledge DistillationLarge Language ModelVision Language ModelMultimodality
🎯 What it does: FeD is proposed, a feedback-guided end-to-end perception-driven strategy utilizing large-scale multimodal language models, achieving complete driving decision-making under single camera input;
🎯 What it does: A few-shot learning method based on diffusion model time steps is proposed, training class-specific low-rank adapters to reconstruct noisy images, thereby achieving debiased classification.
🎯 What it does: A Path eXclusion (PX) method based on a path perspective is proposed, which retains the paths that have the greatest impact on training dynamics by calculating the upper bound of the Neural Tangent Kernel (NTK) trajectory, thus achieving model compression at high sparsity.
🎯 What it does: This paper proposes a FinePOSE framework based on diffusion models, utilizing a fine-grained prompt-driven denoiser to achieve 3D human pose estimation for both single and multiple persons.
🎯 What it does: A multi-person basketball video dataset FineSports covering 10,000 NBA games has been constructed, and a prompt-based spatiotemporal action localization method called PoSTAL has been proposed.
Finsler-Laplace-Beltrami Operators with Application to Shape Analysis
Simon Weber (Technical University of Munich), Daniel Cremers (Technical University of Munich)
CodeSegmentationConvolutional Neural NetworkMesh
🎯 What it does: This paper proposes a Finsler-Laplace-Beltrami operator (FLBO) based on Finsler manifolds and validates its effectiveness in shape matching tasks.
🎯 What it does: In response to the high evaluation cost of text-to-image diffusion models, this paper proposes the FlashEval method, which selects a representative subset from the original prompt set through iterative search and frequency filtering, significantly reducing evaluation time while maintaining high ranking relevance.
Flatten Long-Range Loss Landscapes for Cross-Domain Few-Shot Learning
Yixiong Zou (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
CodeDomain AdaptationMeta LearningImage
🎯 What it does: This paper proposes a method to achieve long-range loss flattening for cross-domain few-shot learning by performing random interpolation on features with different normalization methods in the representation space, thereby enhancing model transfer and fine-tuning performance.
🎯 What it does: This study investigates the parent bias problem in hierarchical semantic segmentation and verifies the advantages of planar classifiers in cross-domain scenarios. It proposes the use of hyperbolic space (Poincaré ball model) to balance inter-class distances, enhancing the accuracy and calibration quality of parent class predictions.
🎯 What it does: A flexible biometric recognition framework (FBR) is proposed, which supports both cross-modal and single-modal recognition of facial, periocular, and soft biometric features simultaneously.
🎯 What it does: Transform the optical flow estimation task into a conditional generation task, using a diffusion model to gradually denoise from random noise to obtain the optical flow field.
🎯 What it does: This paper proposes FlowIE, an efficient image enhancement framework based on rectified flow, which utilizes the generative prior of a pre-trained diffusion model to construct a direct path from noise to clear images, achieving various enhancement tasks in less than 5 steps.
🎯 What it does: This paper proposes an improved interactive segmentation framework called FocSAM, which incorporates dynamic window multi-head self-attention and pixel-level dynamic ReLU based on SAM, achieving a focus on target objects and deeper integration of interactive information.