European Conference on Computer Vision Β· 980 papers
Dense Multimodal Alignment for Open-Vocabulary 3D Scene Understanding
Ruihuang Li (Hong Kong Polytechnic University), Lei Zhang (Joins Hopkins University)
CodeSegmentationAutonomous DrivingConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityPoint Cloud
π― What it does: Propose the Dense Multimodal Alignment (DMA) framework, achieving open-vocabulary 3D scene understanding by establishing dense correspondences among 3D points, 2D pixels, and text.
π― What it does: Proposed the deep perceptual blind image decomposition network DeBNet for restoring clear images in real-world scenarios with mixed adverse weather.
π― What it does: Propose the DetailSemNet model, achieving offline signature verification through local structural matching and detail semantic fusion.
π― What it does: This paper proposes the Detecting As Labeling (DAL) framework, which re-examines the fundamental principles of LiDAR-Camera fusion in 3D object detection and constructs a concise network that performs regression using only point cloud features.
π― What it does: This study proposes DEVIAS, an end-to-end framework based on a decomposition encoder-decoder (Slot Attention + Action Mask Decoder), to learn disentangled representations of actions and scenes in videos.
π― What it does: Proposes DG-PIC, a multi-domain and multi-task point cloud understanding framework that achieves domain generalization through dual-layer feature shifting during testing without requiring model updates.
DGR-MIL: Exploring Diverse Global Representation in Multiple Instance Learning for Whole Slide Image Classification
Wenhui Zhu (Arizona State University), Yalin Wang (Arizona State University)
CodeClassificationTransformerContrastive LearningBiomedical Data
π― What it does: This paper proposes a global diversity aggregation method called DGR-MIL based on multi-instance learning, which models the diversity between WSI instances using a learnable global vector and cross-attention mechanism;
DHR: Dual Features-Driven Hierarchical Rebalancing in Inter- and Intra-Class Regions for Weakly-Supervised Semantic Segmentation
Sanghyun Jo (OGQ), Kyungsu Kim (Massachusetts General Hospital and Harvard Medical School)
CodeSegmentationImageBenchmark
π― What it does: Designed and implemented the DHR (Dual Features-Driven Hierarchical Rebalancing) method, which utilizes unsupervised features (USS) and weakly supervised features (WSS) for hierarchical rebalancing to recover the minority classes overlooked in weakly supervised semantic segmentation.
π― What it does: This paper proposes a "Diagnosing & Re-learning" strategy, which diagnoses the separability of single-modal representation spaces for each modality and subsequently performs soft re-initialization of the corresponding encoder to achieve balanced and enhanced modal learning.
π― What it does: Propose a diffusion model-based matching matrix iterative optimization framework, Diff-Reg, which generates high-quality correspondences by utilizing forward noise diffusion and backward denoising iteration in the doubly stochastic matrix space, further applied to 3D and 2D-3D registration;
DiffBIR: Toward Blind Image Restoration with Generative Diffusion Prior
Xinqi Lin (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Chao Dong (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
CodeRestorationDiffusion modelAuto EncoderImage
π― What it does: Propose a two-stage blind image restoration framework called DiffBIR, which first removes degradation and then reconstructs missing information using a generative diffusion prior.
π― What it does: Propose a new symmetric differentiable Chamfer distance loss, DiffCD, for fitting neural implicit surfaces from sparse noisy point clouds.
π― What it does: Propose DIFFender, which utilizes a pre-trained text-guided diffusion model to locate and recover adversarial patches, forming a unified defense framework.
DiffFAS: Face Anti-Spoofing via Generative Diffusion Models
Xinxu Ge (Tianjin University), Heikki KΓ€lviΓ€inen (Lappeenranta-Lahti University of Technology LUT)
CodeAnomaly DetectionSafty and PrivacyConvolutional Neural NetworkDiffusion modelImageBenchmark
π― What it does: Achieve high-fidelity generation from live faces to spoof faces using diffusion models, and enhance cross-domain and cross-attack facial anti-spoofing performance by incorporating image quality priors.
π― What it does: This paper proposes a new diffusion model based on Vision Transformer called DiffiT, achieving high-quality image generation in both latent and image spaces.
π― What it does: Propose DiffPMAE, a self-supervised point cloud reconstruction architecture that combines Masked Autoencoder with Diffusion Model, applicable for compression, up-sampling, and completion tasks.
π― What it does: Propose a point cloud denoising method based on the SchrΓΆdinger bridge (P2P-Bridge), treating denoising as a reversible data-to-data diffusion process from noisy point clouds to clean point clouds.
π― What it does: Developed a multi-step iterative matting framework called DiffMatte based on a pixel-level denoising diffusion model, which can further refine the alpha matte on top of existing matting encoders.
π― What it does: Propose a pixel-level anomaly detection method based on diffusion models called DOoD, and construct a diverse ADE-OoD benchmark dataset.
π― What it does: This paper proposes an offline reinforcement learning method called Trajectory Diffuser, which significantly improves sampling efficiency while maintaining or enhancing sampling quality by splitting the diffusion model's sampling process into two steps: 'generating feasible trajectories' and 'trajectory optimization'.
π― What it does: Propose a bootstrap variational inference method (DAVI) based on diffusion model priors, which can directly map measurements to implicit posterior distributions under single-step inference to solve noisy inverse problems.
π― What it does: Propose a training-agnostic prompt interpolation noise correction method to correct inverse diffusion starting errors in image-to-image translation with diffusion models, preserving the background structure while precisely editing the target region.
Diffusion-Driven Data Replay: A Novel Approach to Combat Forgetting in Federated Class Continual Learning
Jinglin Liang (South China University of Technology), Qiang Yang (Hong Kong University of Science and Technology)
CodeClassificationData SynthesisFederated LearningSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkDiffusion modelContrastive LearningImage
π― What it does: This paper proposes a federated continual learning framework named DDDR based on diffusion models, aimed at mitigating the catastrophic forgetting problem in federated classification continual learning.
Sung-Hoon Yoon (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
CodeSegmentationTransformerDiffusion modelImage
π― What it does: This paper proposes a weakly supervised semantic segmentation framework that combines diffusion models with vision transformers, enhancing CAM quality through Local Fusion Cross-Attention (LFCA) and Patch Affinity Consistency (PAC).
Diffusion-Refined VQA Annotations for Semi-Supervised Gaze Following
Qiaomu Miao (Stony Brook University), Dimitris Samaras (University of Adelaide)
CodeTransformerVision Language ModelDiffusion modelImageVideoBenchmark
π― What it does: Propose a semi-supervised gaze tracking method that generates Grad-CAM heatmaps using a pre-trained vision-language (VQA) model, denoises and refines them with a diffusion model, and finally produces high-quality pseudo-labels for training the gaze tracking network.
π― What it does: Propose a few-shot handwriting generation method called DiffusionPen based on latent diffusion models, which can reproduce the writing styles of known and unknown writers using only 5 samples and generate highly readable word images.
π― What it does: Propose a one-stage direct knowledge distillation method called 4Ds, which utilizes a teacher network pre-trained on the source domain to directly train a small student network on the target domain without requiring source data.
π― What it does: Built a mobile manipulation framework DISCO based on differentiable scene semantic representation and a two-layer coarse-to-fine control, which can complete navigation and interaction according to verb-noun pairs in commands.
DiscoMatch: Fast Discrete Optimisation for Geometrically Consistent 3D Shape Matching
Paul Roetzer (University of Bonn), Paul Swoboda (Heinrich-Heine University DΓΌsseldorf)
CodeOptimizationMesh
π― What it does: Proposed a fast discrete optimization framework called DiscoMatch, achieving geometrically consistent 3D shape matching through integer linear programming and pre-trained features.
Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discovery
Sukrut Rao (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
CodeExplainability and InterpretabilityRepresentation LearningVision Language ModelAuto EncoderImageTextMultimodality
π― What it does: Propose a reverse concept bottleneck model (DN-CBM), which first automatically discovers interpretable concepts from CLIP features using a sparse autoencoder, then matches text embeddings to name these concepts, and finally constructs an interpretable classifier by using these named concepts as a bottleneck layer.
π― What it does: This paper proposes a Disentangled Masked AutoEncoder (DisMAE), which learns domain-invariant semantic features and domain-specific variation features through a dual-branch architecture to achieve unsupervised domain generalization.
Dissecting Dissonance: Benchmarking Large Multimodal Models Against Self-Contradictory Instructions
Jin Gao (Shanghai Jiao Tong University), Dequan Wang (Shanghai Jiao Tong University)
CodeTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed the Self-Contradictory Instructions (SCI) benchmark, constructed a dataset of 20K self-contradictory instructions, and evaluated the conflict detection capabilities of various large multimodal models.
π― What it does: Propose an unsupervised plugin module called MemE, which can automatically filter distractors in images during NeRF and 3D Gaussian Splatting training, achieving noise-free novel view synthesis.
Distractors-Immune Representation Learning with Cross-modal Contrastive Regularization for Change Captioning
Yunbin Tu (University of Chinese Academy of Sciences), Qingming Huang (Hangzhou Dianzi University)
CodeGenerationRepresentation LearningTransformerVision Language ModelContrastive LearningImageText
π― What it does: This paper studies the visual change description task, proposing an interference-immune representation learning network (DIRL) and cross-modal contrast regularization (CCR). It enhances the robustness of image representations against interferences such as viewpoint and illumination through self-supervised channel correlation/disassociation mechanisms, and generates accurate change descriptions using Transformers.
π― What it does: Proposed a new distribution-aware robust learning framework called DaSC, which can simultaneously address long-tailed distributions and noisy label problems.
Distributionally Robust Loss for Long-Tailed Multi-Label Image Classification
Dekun Lin (Chengdu Institute of Computer Applications, Chinese Academy of Sciences), Xiaolin Qin (Chengdu Institute of Computer Applications, Chinese Academy of Sciences)
π― What it does: Proposed a distribution-robust loss (DR Loss) for long-tailed multi-label image classification, enhancing model robustness on long-tailed data through class-level computation of LSEP loss (C-LSEP) and incorporating a negative gradient constraint (NGC).
π― What it does: This paper proposes a generalization method for 3D object detection from a single source domain. The core idea is to enhance the model's robustness to varying point cloud densities through physical constraint-based point cloud density resampling (PDDA), and to incorporate a self-supervised point cloud densification task into the detection framework, achieving multi-task learning during training and lightweight model adaptation during inference.
Domain Reduction Strategy for Non-Line-of-Sight Imaging
Hyunbo Shim (Yonsei University), Seon Joo Kim (Yonsei University)
CodeOptimizationComputational EfficiencyImagePhysics Related
π― What it does: Proposes an optimization method based on domain reduction that can rapidly reconstruct albedo and surface normals of non-line-of-sight (NLOS) scenes under sparse scanning.
π― What it does: Propose a domain adaptive video deblurring framework based on test-time fuzziness, which utilizes generated pseudo-sharp points and synthetic blurred images to self-optimize the deblurring model, thereby enhancing the restoration performance on target domain videos.
π― What it does: For breast ultrasound image segmentation, the SFRecSAM model based on SAM is proposed, achieving more accurate segmentation through the introduction of spatial-frequency domain feature fusion and dual error correction.
π― What it does: Developed a zero-training, zero-fine-tuning pipeline named DreamDrone, which can generate infinitely long flight perspective sequences along any user-defined camera trajectory from a single RGBD image and text prompts, without constructing 3D point clouds.
π― What it does: Proposes DreamSampler, a unified framework that combines reverse diffusion sampling and score distillation for image editing, restoration, vectorization, and inverse problem solving.
DreamStruct: Understanding Slides and User Interfaces via Synthetic Data Generation
Yi-Hao Peng (Carnegie Mellon University), Amy Pavel (University of Texas Austin)
CodeClassificationRecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes a synthetic structured visual data generation method based on large language models (LLMs) to generate code and render visual data for creating annotated data to build slides and user interfaces.
DriveLM: Driving with Graph Visual Question Answering
Chonghao Sima (Shanghai AI Lab), Hongyang Li (Shanghai AI Lab)
CodeAutonomous DrivingTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityGraphChain-of-Thought
π― What it does: Proposed the DriveLM task, i.e., end-to-end autonomous driving based on graph-structured visual question answering (GVQA), and constructed two large-scale graph question-answering datasets with logical dependencies: DriveLM-nuScenes and DriveLM-CARLA. Corresponding evaluation metrics were provided, and the DriveLM-Agent VLM baseline model was proposed based on this.
π― What it does: This paper proposes a vision parameter-efficient fine-tuning framework called DMLoRA based on dynamic training structures, which enhances model robustness and performance by leveraging multi-branch low-rank adaptation and phased scale learning.
π― What it does: Proposed a self-supervised pre-training framework combined with Cut-Mix data augmentation based on Distortion Sensitivity Map (DSM), dynamically assigning mixed labels through DSM and introducing semantic features via knowledge distillation to achieve no-reference image quality assessment;
π― What it does: Proposed the dual-camera smooth zoom (DCSZ) task, and constructed a virtual camera 'data factory' to generate synthetic training data, thereby fine-tuning existing frame interpolation models to achieve smooth preview.
Dual-stage Hyperspectral Image Classification Model with Spectral Supertoken
Peifu Liu (Beijing Institute of Technology), Jianan Li (Beijing Institute of Technology)
CodeClassificationTransformerImage
π― What it does: Proposed the Dual-stage Spectral Supertoken Classifier (DSTC), which forms spectral supertokens through clustering of spectral derivative features and utilizes Transformers for classification to achieve pixel-level high-precision remote sensing image classification.
π― What it does: Propose the DualBEV framework, unifying 3D-2D and 2D-3D view transformations, achieving BEV feature extraction through dual-perspective probabilistic correspondence.
DVLO: Deep Visual-LiDAR Odometry with Local-to-Global Feature Fusion and Bi-Directional Structure Alignment
Jiuming Liu (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
CodePose EstimationAutonomous DrivingSimultaneous Localization and MappingImageMultimodalityPoint CloudBenchmark
π― What it does: Propose a visual-LiDAR odometry network DVLO based on local-to-global fusion, achieving efficient and fine-grained multimodal feature fusion through bidirectional structural alignment.
π― What it does: Proposed a temporal action detection framework DyFADet based on dynamic feature aggregation (DFA), addressing the issues of insufficient feature discriminability and poor compatibility of the detection head in traditional models.
π― What it does: Propose a dynamic guided adversarial distillation (DGAD) framework that leverages three mechanisms: misclassification-aware partitioning (MAP), error-corrected label exchange (ELS), and prediction consistency regularization (PCR) to enhance the accuracy and robustness of student models on natural images and adversarial examples.
π― What it does: This paper proposes a dynamic retraining-updating Mean Teacher framework for source-free unsupervised object detection (SFOD) tasks.
π― What it does: The delayed Ξ΅-shrink method is proposed within the once-for-all framework, which first partially preheats the full model, then gradually introduces subnetworks, and efficiently trains with shared weights through Ξ΅-Shrinking learning rate scheduling and IKD-Warmup.
E3M: Zero-Shot Spatio-Temporal Video Grounding with Expectation-Maximization Multimodal Modulation
Peijun Bao (Nanyang Technological University), Alex Kot
CodeRetrievalVision Language ModelContrastive LearningVideoTextMultimodalityBenchmark
π― What it does: Propose a zero-shot spatiotemporal video localization method called E3M, which directly locates target objects in videos during testing using a pre-trained CLIP model.
π― What it does: This paper first constructs the E3V-K5 dataset of motion video energy expenditure based on COSMED K5 ground-truth measurements, and proposes the E3SFormer model, which simultaneously performs action recognition and energy regression using human skeletal videos to estimate energy expenditure at the video level.
π― What it does: Propose a new classifier pre-tuning method called NeST, which helps better initialize classifiers in new tasks for class-incremental semantic segmentation
Echoes of the Past: Boosting Long-tail Recognition via Reflective Learning
Qihao Zhao (Beijing University of Chemical Technology), Jun Liu (Singapore University of Technology and Design)
CodeClassificationRecognitionImageBenchmark
π― What it does: Propose a reflection learning framework that improves long-tailed image recognition performance through three stages: review, summarize, and correct.
Efficient Bias Mitigation Without Privileged Information
Mateo Espinosa Zarlenga (University of Cambridge), Alice Xiang (Sony AI)
CodeClassificationDomain AdaptationImage
π― What it does: Proposes a bias mitigation framework named TAB without group information, which partitions samples and generates group-balanced datasets by leveraging the complete training history of an auxiliary model, followed by retraining a robust model from this dataset.
π― What it does: Discover redundancy in query-key interactions within diffusion Transformers, proposing the use of mediator tokens to compress attention and dynamically adjust the number of mediator tokens based on denoising steps, achieving an efficient diffusion Transformer with linear complexity;
π― What it does: This paper proposes Decorruptor, which fine-tunes the Latent Diffusion Model using an improved corruption modeling scheme, enabling the generation of clear images by editing corrupted images during testing and accelerating inference.
π― What it does: By freezing the CLIP vision and text encoders, we achieve efficient few-shot action recognition through multi-level post-reasoning and interactive spatiotemporal reasoning.
π― What it does: This paper proposes FADformer, a frequency-domain based Transformer framework, which achieves efficient single-image deraining using a frequency-domain convolution mixer and prior-gated feed-forward networks.
Efficient Inference of Vision Instruction-Following Models with Elastic Cache
Zuyan Liu (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeComputational EfficiencyTransformerVision Language ModelMultimodality
π― What it does: Propose the Elastic Cache method, which compresses and accelerates the KV cache during inference for large-scale vision instruction following models (e.g., LLaVA, Qwen-VL). By adopting importance-driven cache merging (anchor-point + bucket merging) during the instruction encoding phase and using fixed-point removal (retaining the initial and latest KV pairs) during the output generation phase, a training-agnostic multi-stage acceleration is achieved.
π― What it does: This paper proposes an automatic method for filtering and replacing ASR transcripts with human-written instructions (Sieve & Swap), and implements step localization and text generation for cooking videos using a pre-trained Procedure Transformer (ProcX).
π― What it does: Proposes an implicit training method for Spiking Neural Networks (SNN) based on a multi-parallel implicit flow (MPIS) architecture, achieving fast convergence while maintaining low latency, low memory consumption, and low sparsity.
EgoCVR: An Egocentric Benchmark for Fine-Grained Composed Video Retrieval
Thomas Hummel (University of TΓΌbingen), Zeynep Akata (TU Munich)
CodeRetrievalLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: This paper proposes a fine-grained first-person perspective video retrieval benchmark called EgoCVR, evaluates multiple vision-language models on this benchmark, and designs a training-agnostic re-ranking framework named TFR-CVR.
π― What it does: This paper introduces the EgoExo-Fitness dataset, which collects synchronized egocentric (front and downward views) and exocentric (front, left-front, right-front) full-body fitness action videos. It provides rich annotations, including two-level time boundaries, technical keypoint validation, natural language comments, and action quality scores. Based on this dataset, five benchmark tasks are constructed: action classification, action localization, cross-perspective sequence verification, cross-perspective skill judgment, and a novel instruction execution verification.
EgoPoser: Robust Real-Time Egocentric Pose Estimation from Sparse and Intermittent Observations Everywhere
Jiaxi Jiang (ETH ZΓΌrich), Christian Holz (ETH ZΓΌrich)
CodePose EstimationComputational EfficiencyTransformerPoint CloudMeshTime Series
π― What it does: To address the sparse and discontinuous head and hand position information provided by head-mounted devices, this paper proposes a real-time full-body pose estimation method called EgoPoser, which can maintain high-accuracy and coherent pose outputs even when hands are out of the field of view.
EINet: Point Cloud Completion via Extrapolation and Interpolation
Pingping Cai (University of South Carolina), Song Wang (University of South Carolina)
CodeRestorationTransformerPoint Cloud
π― What it does: Propose a new point cloud completion framework called EINet, which uses extrapolation in the feature space to complete missing shapes and interpolation in the feature space to upsample point clouds.
Elegantly Written: Disentangling Writer and Character Styles for Enhancing Online Chinese Handwriting
Yu Liu (University Putra Malaysia), Cunrui Wang (Dalian Minzu University)
CodeGenerationTransformerSequential
π― What it does: This paper proposes an online Chinese handwriting trajectory beautification method based on a sequence Transformer, which can learn writing styles from a small number of user samples and optimize and beautify the handwriting trajectories while preserving the original text content.
π― What it does: Designed and proposed a pluggable disambiguation network, AENet, to eliminate feature ambiguity in few-shot segmentation, enhancing the matching quality between query and support foreground features, thereby significantly improving the segmentation performance of multiple baseline models.
π― What it does: Proposes an online unsupervised video stitching and stabilization framework named StabStitch, specifically addressing the 'warping shake' problem in video stitching;
Elucidating the Hierarchical Nature of Behavior with Masked Autoencoders
Lucas Stoffl (Ecole Polytechnique Federale de Lausanne), Alexander Mathis (Ecole Polytechnique Federale de Lausanne)
CodeExplainability and InterpretabilityRepresentation LearningTransformerAuto EncoderVideoTime SeriesBenchmark
π― What it does: This paper proposes a hierarchical masked autoencoder called hBehaveMAE, which is validated on a newly constructed synthetic basketball dataset Shot7M2 and the extended human action benchmark hBABEL.
Elysium: Exploring Object-level Perception in Videos through Semantic Integration Using MLLMs
Han Wang (Bytedance Inc), Can Huang (Bytedance Inc)
CodeObject TrackingTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
π― What it does: Constructed and released a large-scale video object perception dataset ElysiumTrack-1M (1.27 million trajectories + descriptions), and proposed an end-to-end multimodal large language model Elysium that can directly complete video-level and object-level tasks; visual token compression is achieved through T-Selector, balancing frame rate and performance.
π― What it does: Proposed a Transformer-based Encoder-Decoder structure called EDAFormer, combining Embedding-Free Attention and Inference Spatial Reduction to achieve efficient semantic segmentation.
Yunsong Zhou (OpenDriveLab at Shanghai AI Lab), Hongyang Li (Shanghai Jiao Tong University)
CodeAutonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodality
π― What it does: This paper proposes the Embodied Language Model (ELM), aiming to achieve a four-dimensional, full-space, long-term temporal embodied understanding of driving scenarios.
π― What it does: Propose a Hierarchical Feature Refinement Network (HFRN) and a Cross-Modal Adaptive Feature Refinement (CAFR) module to integrate features from event cameras and traditional frame cameras, improving target detection performance in complex environments.
π― What it does: This paper addresses the inefficiency of Mask Token in MIM pre-training by proposing the Masked Token Optimization (MTO) method, which enhances pre-training efficiency by analyzing and optimizing the learning process of mask tokens in Transformers.
π― What it does: Proposes a new knowledge transfer paradigm called Knowledge in One Prompt (KiOP), which encapsulates knowledge from multiple models into a single visual prompt without modifying the source models or accessing the original training data.
π― What it does: Propose a multi-modal MRI fusion method based on energy models, Energy-induced Explicit Propagation and Alignment (EPA), which explicitly quantifies and optimizes aggregation patterns under different diseases.
π― What it does: Enhance network sparsification through self-distillation training, proposing the STP framework to achieve one-stage multi-dimensional (depth + width) structured pruning. Before pruning, self-distillation is used to maintain the magnitude of pruned weights, enhancing the expressiveness of retained weights.
π― What it does: Fine-tune the text encoder using reinforcement learning combined with LoRA to enhance text-image alignment and visual quality in diffusion models, and seamlessly integrate with existing U-Net fine-tuning models;
π― What it does: This paper proposes the Diode optimizer specifically designed for binary neural networks (BNNs), achieving random sign descent updates without hidden weights by leveraging gradient sign information.
π― What it does: This paper proposes the StableVSR method based on diffusion models for video super-resolution, aiming to enhance perceptual quality and ensure temporal consistency.
Enhancing Plausibility Evaluation for Generated Designs with Denoising Autoencoder
Jiajie Fan (BMW Group), Hao Wang (Leiden University)
CodeGenerationAuto EncoderImageMesh
π― What it does: Proposed a structural feasibility assessment metric called FDD based on denoising autoencoders to evaluate the structural rationality of generated design images, and validated its superiority on multiple datasets.
Enhancing Recipe Retrieval with Foundation Models: A Data Augmentation Perspective
Fangzhou Song (University of Science and Technology of China), Shuo Wang (University of Science and Technology of China)
CodeSegmentationRetrievalData-Centric LearningTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBenchmark
π― What it does: Propose to use Llama2 and SAM for data augmentation, and insert a lightweight adapter and multi-level circular loss into CLIP to enhance cross-modal recipe retrieval performance.
Enhancing Semantic Fidelity in Text-to-Image Synthesis: Attention Regulation in Diffusion Models
Yang Zhang (National University of Singapore), Kenji Kawaguchi (National University of Singapore)
CodeGenerationVision Language ModelDiffusion modelImageText
π― What it does: This paper proposes a training-agnostic method that enables real-time modulation of cross-attention during the inference phase of diffusion models, aiming to enhance semantic fidelity in text-to-image synthesis.
π― What it does: Propose a low-confidence pseudo-label distillation loss to help source-free domain adaptive object detection models identify overlooked hard-to-detect instances.