🎯 What it does: WaffleIron is proposed, a novel 3D backbone network that achieves semantic segmentation of vehicle-mounted LiDAR point clouds using only MLP and dense 2D convolutions, without the need for sparse convolutions.
🎯 What it does: Proposes the µ Split method, which efficiently performs image decomposition in fluorescence microscopy images using lateral context (LC), splitting the superimposed two-channel images into their respective channels.
VAD: Vectorized Scene Representation for Efficient Autonomous Driving
Bo Jiang (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
CodeAutonomous DrivingTransformerMultimodality
🎯 What it does: This paper proposes an end-to-end autonomous driving framework called VAD, which is completely based on vectorized scene representation (vectorized maps and vectorized motion) and directly outputs driving trajectories from multi-view cameras, eliminating the traditional rasterized perception and post-processing steps.
🎯 What it does: VADER is a set of spatiotemporal matching, alignment, and differential techniques for video clips, capable of retrieving original videos from massive video libraries and locating segments, followed by visualizing and annotating edited areas;
🎯 What it does: An unsupervised perspective representation learning and perspective-aware 3D point cloud completion network, VAPCNet, is proposed, capable of completing sparse point clouds from unknown viewpoints.
Variational Degeneration to Structural Refinement: A Unified Framework for Superimposed Image Decomposition
Wenyu Li (Tianjin University), Yue Lang (Hebei University of Technology)
CodeRestorationAuto EncoderImage
🎯 What it does: A unified framework VDSR is proposed for the decomposition of single mixed images, and it is extended to tasks such as rain removal, reflection removal, and shadow removal.
🎯 What it does: This paper proposes Versatile Diffusion (VD), a multi-stream multi-modal diffusion model capable of performing tasks such as text-to-image, image-to-text, and image variant generation within the same network, and supports dual/multi-context mixing and style/semantic unsupervised decoupling.
CodeSafty and PrivacyAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: A privacy leakage attack on graph neural networks called VertexSerum is proposed, which amplifies connection information and successfully steals the link relationships between nodes by applying lightweight adversarial perturbations to the node features during the training phase.
VI-Net: Boosting Category-level 6D Object Pose Estimation via Learning Decoupled Rotations on the Spherical Representations
Jiehong Lin (South China University of Technology), Kui Jia (South China University of Technology)
CodeObject DetectionPose EstimationPoint Cloud
🎯 What it does: VI-Net is proposed to achieve high-precision category-level 6D object pose estimation by performing feature learning on the sphere and decoupling rotation into viewpoint rotation and in-plane rotation, utilizing V-Branch and I-Branch for binary classification and regression, respectively.
🎯 What it does: A unified video multi-bad weather removal framework, ViWS-Net, has been designed to restore video frames contaminated by various weather conditions such as rain, fog, and snow all at once.
🎯 What it does: This paper proposes an auxiliary training framework VOS-VFI based on Video Object Segmentation (VOS) to enhance the clarity and accuracy of foreground object boundaries in Video Frame Interpolation (VFI) models.
Federica Arrigoni (Politecnico di Milano), Andrea Fusiello (University of Udine)
CodeGraph
🎯 What it does: This paper studies the view graph solvability problem in structured light 3D reconstruction and proposes an efficient method to test finite solvability in large-scale uncalibrated graphs and extract the maximum solvable subgraph.
🎯 What it does: This study investigates whether VLM trained on high 'pairwise complexity' datasets can learn fine-grained region-attribute relationships, and proposes ViLLA, which retrains VLM by generating region-attribute pairs through self-supervised mapping.
🎯 What it does: A posture-clothing keypoint guided virtual try-on method KGI has been designed and implemented, which can generate high-fidelity try-on effects that maintain the shape and pattern of clothing given a portrait and clothing image.
Visible-Infrared Person Re-Identification via Semantic Alignment and Affinity Inference
Xingye Fang (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
CodeRecognitionRetrievalImageMultimodality
🎯 What it does: This paper proposes an end-to-end visible-infrared person re-identification framework called SAAI, which combines semantic alignment feature learning and an affinity reasoning module to achieve cross-modal person matching.
Vision Grid Transformer for Document Layout Analysis
Cheng Da (Alibaba Group), Cong Yao (Alibaba Group)
CodeObject DetectionSegmentationTransformerImage
🎯 What it does: This paper proposes a Dual-Stream Vision Grid Transformer (VGT), which utilizes the Grid Transformer (GiT) for token-level and segment-level semantic pre-training on a two-dimensional document grid, and then combines it with ViT to perform document layout analysis.
Vision Relation Transformer for Unbiased Scene Graph Generation
Gopika Sudhakaran (Technical University of Darmstadt), Stefan Roth (Technical University of Darmstadt)
CodeClassificationObject DetectionGenerationTransformerMixture of ExpertsImageMultimodality
🎯 What it does: This paper proposes a Vision Relation Transformer (VETO) that enhances relation prediction through local-level entity patch generation and multimodal fusion, combined with a Mutually Exclusive ExperT (MEET) multi-expert learning strategy to achieve unbiased scene graph generation.
CodeSegmentationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: A general visual model interpretation method called Iterated Integrated Attributions (IIA) is proposed, which generates precise explanation heatmaps through iterative integration of inputs, internal representations, and their gradients.
CodeObject DetectionGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an automated data augmentation framework called CaCao, which extracts fine-grained predicates from large-scale pre-trained language models using a visual prompting language model to address the long-tail distribution problem in visual relation generation. It further introduces the Epic open-world predicate generation module to achieve zero-shot predicate prediction.
VL-PET: Vision-and-Language Parameter-Efficient Tuning via Granularity Control
Zi-Yuan Hu (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
CodeTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: A parameter-efficient fine-tuning framework for visual-language tasks, VL-PET, is proposed, which improves the adaptation effect of PLM through a granularity control mechanism, lightweight module design, and multi-head modular modifications.
VLN-PETL: Parameter-Efficient Transfer Learning for Vision-and-Language Navigation
Yanyuan Qiao (Australian Institute for Machine Learning, University of Adelaide), Qi Wu (Australian Institute for Machine Learning, University of Adelaide)
CodeTransformerPrompt EngineeringMultimodality
🎯 What it does: The study applies the Parameter-Efficient Transfer Learning (PETL) method to Visual Language Navigation (VLN) and proposes the VLN-PETL framework, designing the Historical Interaction Boost (HIB) and Cross-Modal Interaction Boost (CIB) modules.
Eric Slyman (Oregon State University), Stefan Lee (Oregon State University)
CodeObject DetectionRecommendation SystemTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper presents an interactive system called VL Slice, which helps researchers quickly discover and construct subsets (slices) that align with the bias dimensions of visual language models within unannotated image collections.
Waffling Around for Performance: Visual Classification with Random Words and Broad Concepts
Karsten Roth (University of Tuebingen), Zeynep Akata (Google DeepMind)
CodeClassificationLarge Language ModelVision Language ModelImage
🎯 What it does: A zero-shot image classification method called Waffle CLIP is proposed, which replaces the fine-grained descriptions originally generated by LLMs with random words or character sequences added to the category prompts; it also proposes using LLMs to automatically extract high-level concepts to alleviate category ambiguity.
🎯 What it does: A dynamically expandable Wasserstein Variational Autoencoder (WEVAE) is proposed, which determines when to add new components using Wasserstein distance and filters memory samples with an energy function to enhance knowledge diversity.
Weakly Supervised Learning of Semantic Correspondence through Cascaded Online Correspondence Refinement
Yiwen Huang (Fudan University), Weifeng Ge (Fudan University)
CodeRecognitionSegmentationTransformerImage
🎯 What it does: A weakly supervised semantic correspondence framework based on multi-instance learning is proposed, which can learn pixel-level correspondence using only image-level labels.
🎯 What it does: A weakly supervised action localization framework AHLM is proposed, which utilizes a hierarchical latent attention model to simultaneously detect feature semantic change points and accurately locate action boundaries.
🎯 What it does: A box-free GAN ownership verification method is designed, which proves copyright solely by detecting generated images using the unique distribution learned by the generator.
🎯 What it does: By using the four simplest arithmetic operations of addition, subtraction, multiplication, and division between frame features to generate auxiliary temporal information and embedding it into the original features, temporal modeling of videos is achieved.
🎯 What it does: This paper proposes a non-local feature aggregation method based on disparity geometric constraints—Epipolar Transformer—to improve feature matching quality and reconstruction accuracy in multi-view stereo reconstruction.
🎯 What it does: A representation calibration method based on contrastive learning (RCAL) is proposed, which first obtains robust features using unsupervised contrastive learning, and then performs distributed calibration and individual calibration on the representation distribution of each category, thereby addressing both noisy labels and long-tail distributions simultaneously.
🎯 What it does: This paper proposes a learnable Adaptive Prompt Generator (APG) that replaces traditional fixed prompt pools by generating task-specific prompts in incremental learning, thereby reducing the semantic gap between pre-training tasks and subsequent tasks, and supporting zero-shot incremental learning.
Why Is Prompt Tuning for Vision-Language Models Robust to Noisy Labels?
Cheng-En Wu (University of Wisconsin-Madison), Linjie Yang (ByteDance Inc.)
CodeClassificationRecognitionTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper studies how prompt tuning of visual-language models (such as CLIP) maintains high performance under noisy labels, demonstrating their inherent robustness to noise and further enhancing unsupervised prompt tuning through the use of Generalized Cross Entropy (GCE) loss and random pseudo-labeling.
🎯 What it does: This paper studies the computational-performance trade-off of Deep Ensembles and single model scaling in uncertainty estimation tasks, and proposes an early exit cascade strategy for decision boundary windows to significantly reduce inference costs while ensuring the quality of uncertainty.
🎯 What it does: A Prototypical Memory Attention (PMA) network is designed, utilizing the activations of past training samples as learnable memory vectors directly embedded in the self-attention layer of the Transformer for image caption generation.
X-VoE: Measuring eXplanatory Violation of Expectation in Physical Events
Bo Dai (Peking University), Yixin Zhu (Peking University)
CodeExplainability and InterpretabilityTransformerAuto EncoderVideoPhysics Related
🎯 What it does: The X-VoE explanatory violation experiment dataset and the XPL model are proposed to evaluate and enhance AI's understanding and explanatory capabilities of intuitive physics.
🎯 What it does: This paper proposes the SeedAL method, which automatically constructs initial annotation seeds for 3D point cloud semantic segmentation, significantly enhancing the effectiveness of active learning.
Zero-Shot Composed Image Retrieval with Textual Inversion
Alberto Baldrati (University of Florence), Alberto Del Bimbo (University of Florence)
CodeRetrievalKnowledge DistillationVision Language ModelContrastive LearningImageText
🎯 What it does: A zero-shot combined image retrieval method called SEARLE is proposed, which maps reference images to pseudo-word tokens using text inversion, and retrieves them in the CLIP text space after concatenating with relative descriptions, completely independent of labeled data.
🎯 What it does: This paper proposes a diffusion model method based on zero-shot contrastive loss (ZeCon) for text-guided image style transfer, which maintains content consistency without additional training or fine-tuning.