arXivSub Start free trial

CVPR 2024 Papers — Page 12

IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2716 papers

Hourglass Tokenizer for Efficient Transformer-Based 3D Human Pose Estimation

Wenhao Li (Peking University), Nicu Sebe (University of Trento)

Pose EstimationComputational EfficiencyTransformerVideo

🎯 What it does: A pluggable pruning and recovery framework called Hourglass Tokenizer (HoT) is proposed to enhance the efficiency of transformer models in video-based 3D joint pose estimation.

HouseCat6D - A Large-Scale Multi-Modal Category Level 6D Object Perception Dataset with Household Objects in Realistic Scenarios

HyunJun Jung (Technical University of Munich), Benjamin Busam (Toyota Motor Europe)

Object DetectionPose EstimationRobotic IntelligenceImageMultimodality

🎯 What it does: This paper constructs the HouseCat6D dataset, providing high-precision 6D pose and grasp annotations for 194 household objects in 41 unmarked, real-world scenes.

How Far Can We Compress Instant-NGP-Based NeRF?

Yihang Chen (Shanghai Jiao Tong University), Jianfei Cai (Monash University)

CompressionNeural Radiance FieldAuto EncoderImage

🎯 What it does: A Context-based NeRF Compression (CNC) framework is proposed, which compresses the multi-resolution hash features of Instant-NGP through a context model, significantly reducing storage requirements.

How to Configure Good In-Context Sequence for Visual Question Answering

Li Li (Southeast University), Xu Yang (Southeast University)

RecognitionRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation

🎯 What it does: This study investigates how to configure the context sequence of large visual language models in the visual question answering task by retrieving and manipulating examples to enhance their few-shot reasoning capabilities.

How to Handle Sketch-Abstraction in Sketch-Based Image Retrieval?

Subhadeep Koley (University of Surrey), Yi-Zhe Song (University of Surrey)

RetrievalGenerative Adversarial NetworkImage

🎯 What it does: Proposes an FG-SBIR framework that utilizes the latent space of StyleGAN to construct a k×d feature matrix and dynamically match the level of abstraction.

How to Make Cross Encoder a Good Teacher for Efficient Image-Text Retrieval?

Yuxin Chen (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)

RetrievalKnowledge DistillationContrastive LearningImageTextMultimodality

🎯 What it does: This paper proposes Contrastive Partial Ranking Distillation (CPRD), which distills the ranking knowledge of hard negative samples from a cross-encoder to a dual-encoder through contrastive learning, thereby improving the accuracy of image-text retrieval.

How to Train Neural Field Representations: A Comprehensive Study and Benchmark

Samuele Papa (University of Amsterdam), Efstratios Gavves (University of Amsterdam)

ClassificationRepresentation LearningGraph Neural NetworkTransformerNeural Radiance FieldImageBenchmark

🎯 What it does: This paper presents the Fit-a-NeF library, which efficiently fits millions of neural fields (NeFs) using the parallelization features of JAX (vmap, JIT), and systematically evaluates their performance as data representations in downstream classification tasks. The authors also constructed the Neural Field Arena benchmark based on experimental results, providing datasets and baseline code for neural field versions such as MNIST, CIFAR-10, MicroImageNet, and ShapeNetv2-10.

HPL-ESS: Hybrid Pseudo-Labeling for Unsupervised Event-based Semantic Segmentation

Linglin Jing (Loughborough University), Xuelong Li (Institute of Artificial Intelligence)

SegmentationDomain AdaptationImage

🎯 What it does: This paper proposes HPL-ESS, a hybrid pseudo-label framework that integrates self-training and offline event-image reconstruction for unsupervised event semantic segmentation.

HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention

Xiaolong Tang (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)

Autonomous DrivingGraph Neural NetworkTransformerMultimodalityTime Series

🎯 What it does: A dynamic trajectory prediction model HPNet based on historical prediction attention is proposed, which can capture the relationships between predictions at consecutive time steps using historical prediction information.

HRVDA: High-Resolution Visual Document Assistant

Chaohu Liu (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)

TransformerLarge Language ModelSupervised Fine-TuningImageText

🎯 What it does: A high-resolution visual document assistant (HRVDA) is proposed, capable of directly understanding high-resolution document images for multiple tasks without the need for OCR preprocessing.

HUGS: Holistic Urban 3D Scene Understanding via Gaussian Splatting

Hongyu Zhou (Zhejiang University), Yiyi Liao (Zhejiang University)

Object DetectionSegmentationAutonomous DrivingGaussian SplattingOptical FlowImagePoint Cloud

🎯 What it does: By using 3D Gaussian splatting, RGB images with pose are elevated to 3D space, achieving joint reconstruction and real-time rendering of geometric, semantic, and motion aspects of urban scenes.

HUGS: Human Gaussian Splats

Muhammed Kocabas (Apple), Anurag Ranjan (Apple)

GenerationData SynthesisPose EstimationGaussian SplattingVideo

🎯 What it does: Utilizing single-camera video (50-100 frames) to train a 3D Gaussian splatting representation of the human body and static scenes, enabling human animation and scene visualization.

Human Gaussian Splatting: Real-time Rendering of Animatable Avatars

Arthur Moreau (Huawei Noah's Ark Lab), Eduardo Pérez-Pellitero (Huawei Noah's Ark Lab)

GenerationPose EstimationGaussian SplattingVideo

🎯 What it does: A 3D Gaussian Splatting-based animatable human head model, HuGS, has been developed, capable of learning from multi-view videos and rendering realistic 3D portraits in real-time under any pose.

Human Motion Prediction Under Unexpected Perturbation

Jiangbei Yue (University of Leeds), He Wang (University College London)

Pose EstimationRobotic IntelligenceRecurrent Neural NetworkTime SeriesSequential

🎯 What it does: A motion prediction task for humans under unexpected disturbances is proposed, along with a corresponding dataset.

HumanGaussian: Text-Driven 3D Human Generation with Gaussian Splatting

Xian Liu (Chinese University of Hong Kong), Ziwei Liu (Chinese University of Hong Kong)

GenerationData SynthesisScore-based ModelGaussian SplattingImageTextMultimodality

🎯 What it does: This paper presents HumanGaussian, which utilizes structure-aware Score Distillation Sampling (SDS) and annealed negative prompting guidance to generate high-quality 3D human models from text within the 3D Gaussian Splatting framework.

HumanNeRF-SE: A Simple yet Effective Approach to Animate HumanNeRF with Diverse Poses

Caoyuan Ma (Wuhan University), Zheng Wang (Wuhan University)

GenerationPose EstimationNeural Radiance FieldVideo

🎯 What it does: This paper proposes a simplified HumanNeRF structure based on the SMPL prior—HumanNeRF-SE, which can be trained with very few monocular video frames and generate high-quality human views in any pose.

HumanNorm: Learning Normal Diffusion Model for High-quality and Realistic 3D Human Generation

Xin Huang (Northwestern Polytechnical University), Qing Wang (Northwestern Polytechnical University)

GenerationData SynthesisDiffusion modelScore-based ModelPoint CloudMesh

🎯 What it does: Generate high-quality and realistic 3D human models from text descriptions, including geometry and texture, and support subsequent animation and editing.

HumanRef: Single Image to 3D Human Generation via Reference-Guided Diffusion

Jingbo Zhang, Jing Liao

GenerationDiffusion modelScore-based ModelImageMesh

🎯 What it does: This paper proposes a method for generating 3D clothed human models from a single reference image, utilizing an implicit SDF network to achieve consistent reconstruction of geometry and texture.

HumMUSS: Human Motion Understanding using State Space Models

Arnab Mondal, Denis Tome

RecognitionPose EstimationVideoMesh

🎯 What it does: A non-attention spatiotemporal state space model called HumMUSS is proposed for predicting 3D poses, human meshes, and action recognition from 2D keypoint videos.

HUNTER: Unsupervised Human-centric 3D Detection via Transferring Knowledge from Synthetic Instances to Real Scenes

Yichen Yao (ShanghaiTech University), Yuexin Ma (Chinese University of Hong Kong)

Object DetectionPose EstimationPoint CloudMesh

🎯 What it does: This paper proposes an unsupervised 3D human detection method called HUNTER, which achieves detection by transferring knowledge from synthetic human models to real point cloud scenes.

Hunting Attributes: Context Prototype-Aware Learning for Weakly Supervised Semantic Segmentation

Feilong Tang (Monash University), Zongyuan Ge (Monash University)

SegmentationConvolutional Neural NetworkImage

🎯 What it does: A Context Prototype‑Aware Learning (CPAL) method is designed and implemented for weakly supervised semantic segmentation, enhancing the integrity and accuracy of CAM through improved category prototype awareness.

Hybrid Functional Maps for Crease-Aware Non-Isometric Shape Matching

Lennart Bastian (Technical University of Munich), Zorah Lähner (University of Siegen)

OptimizationDiffusion modelPoint CloudMeshBenchmark

🎯 What it does: A mixed spectral basis obtained from the Laplace-Beltrami operator and the elastic shell Hessian operator is proposed to achieve non-isometric shape matching within the functional mapping framework.

Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective

Jinjing Zhao (University of Sydney), Chang Xu (University of Sydney)

Object DetectionData-Centric LearningTransformerImage

🎯 What it does: This paper views the DETR series as an improvement of the RPN-refiner in Faster R-CNN, systematically analyzing the key improvement factors of DETR and proposing the Hybrid Proposal Refiner (HPR), which integrates various proposal refinement modules into a unified framework.

HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces

Haithem Turki (Meta Reality Labs), Christian Richardt (Meta Reality Labs)

GenerationComputational EfficiencyNeural Radiance FieldGaussian SplattingMesh

🎯 What it does: We propose HybridNeRF, a hybrid neural field representation that integrates surface and volume rendering for high-quality and real-time VR-level view synthesis.

Hyper-MD: Mesh Denoising with Customized Parameters Aware of Noise Intensity and Geometric Characteristics

Xingtao Wang (Harbin Institute of Technology), Debin Zhao (Harbin Institute of Technology)

RestorationGraph Neural NetworkMesh

🎯 What it does: A hyper-network-based mesh denoising method called Hyper-MD is proposed, which can adaptively generate denoising parameters for each triangular face and output denoised normals.

Hyperbolic Anomaly Detection

Huimin Li (Beihang University), Junlin Hu (Beihang University)

Anomaly DetectionKnowledge DistillationConvolutional Neural NetworkTransformerContrastive LearningImage

🎯 What it does: A self-supervised method for industrial image defect detection in hyperbolic space, called HypAD, is proposed, which optimizes feature embedding using hyperbolic distance to achieve pixel-level anomaly localization.

Hyperbolic Learning with Synthetic Captions for Open-World Detection

Fanjie Kong (Duke University), Davide Modolo (Amazon Web Services AI Labs)

Object DetectionTransformerVision Language ModelContrastive LearningImageTextMultimodality

🎯 What it does: This paper proposes a method for generating synthetic subtitles using a pre-trained vision-language model for open-world object detection; it aligns visual features with subtitle embeddings through a hierarchical hyperbolic structure, thereby enhancing the model's generalization ability to new categories and free-text descriptions.

HyperDreamBooth: HyperNetworks for Fast Personalization of Text-to-Image Models

Nataniel Ruiz, Kfir Aberman

RecognitionGenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImage

🎯 What it does: Using a single face image, quickly generate personalized diffusion model weights with a hypernetwork, and achieve high-fidelity diverse generation through fine-tuning.

HyperSDFusion: Bridging Hierarchical Structures in Language and Geometry for Enhanced 3D Text2Shape Generation

Zhiying Leng (Beihang University), Federico Tombari (Technical University of Munich)

GenerationData SynthesisGraph Neural NetworkDiffusion modelTextMultimodality

🎯 What it does: A HyperSDFusion based on a dual-branch diffusion model is proposed, which can generate three-dimensional shapes from natural language text.

Hyperspherical Classification with Dynamic Label-to-Prototype Assignment

Mohammad Saeed Ebrahimi Saadabadi (West Virginia University), Nasser M. Nasrabadi (West Virginia University)

ClassificationImage

🎯 What it does: This paper proposes a non-parametric classification framework: fixing prototypes that are evenly distributed on the sphere, and achieving classification through dynamic label-to-prototype assignment during the training process, thereby fully utilizing the metric space and capturing the interrelationships between categories.

I'M HOI: Inertia-aware Monocular Capture of 3D Human-Object Interactions

Chengfeng Zhao (ShanghaiTech University), Lan Xu (ShanghaiTech University)

Object DetectionPose EstimationConvolutional Neural NetworkDiffusion modelImageVideoMesh

🎯 What it does: This paper achieves real-time capture of 3D human and object motion in human-computer interaction scenarios by combining a single camera with an object-attached IMU sensor.

IBD-SLAM: Learning Image-Based Depth Fusion for Generalizable SLAM

Minghao Yin (University of Hong Kong), Kai Han (University of Hong Kong)

Depth EstimationOptimizationNeural Radiance FieldSimultaneous Localization and MappingImage

🎯 What it does: A transferable image-based depth fusion framework IBD-SLAM is proposed, which utilizes NeRF to represent scenes and directly performs 3D reconstruction and visual SLAM on unseen scenes without the need for optimization for each scene.

ICON: Incremental CONfidence for Joint Pose and Radiance Field Optimization

Weiyao Wang (Meta), Matt Feiszli (Meta)

Pose EstimationOptimizationNeural Radiance FieldSimultaneous Localization and MappingPoint Cloud

🎯 What it does: This paper proposes a neural radiance field training method called ICON, which uses incremental registration and adaptive confidence optimization without prior pose information.

ICP-Flow: LiDAR Scene Flow Estimation with ICP

Yancong Lin (Delft University of Technology), Holger Caesar (Delft University of Technology)

Autonomous DrivingOptimizationSimultaneous Localization and MappingOptical FlowPoint Cloud

🎯 What it does: This paper proposes an unsupervised LiDAR scene flow estimation method called ICP-Flow, which is completely based on ICP. It segments the point cloud into clusters using the rigid body motion assumption, registers the clusters, and recovers point-level scene flow through inter-cluster rigid transformations.

ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation

Jia-Hao Wu (National Yang Ming Chiao Tung University), Yen-Yu Lin (National Yang Ming Chiao Tung University)

RestorationGenerationTransformerDiffusion modelOptical FlowImage

🎯 What it does: This paper proposes an implicit diffusion-based deblurring method called ID-Blau, which utilizes pixel-level continuous blur condition maps (direction and magnitude) to transform sharp images into controllable blurred images, thereby achieving diversified data augmentation.

ID-like Prompt Learning for Few-Shot Out-of-Distribution Detection

Yichen Bai (Tianjin University), Changqing Zhang (Tianjin University)

Anomaly DetectionPrompt EngineeringVision Language ModelContrastive LearningImage

🎯 What it does: This work proposes a CLIP-based ID-like Prompt learning framework for detecting out-of-distribution (OOD) samples in few-shot scenarios.

Identifying Important Group of Pixels using Interactions

Kosuke Sumiyasu (Chiba University), Hiroshi Kera (Chiba University)

ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerImage

🎯 What it does: A pixel visualization method based on game-theoretic interaction, MoXI, is proposed to efficiently identify the set of pixels that most significantly affect the confidence of image classifiers.

IDGuard: Robust General Identity-centric POI Proactive Defense Against Face Editing Abuse

Yunshu Dai, Fangjun Huang

Convolutional Neural NetworkImage

🎯 What it does: This paper explores a specific problem in the field of computer vision and proposes a new solution.

IIRP-Net: Iterative Inference Residual Pyramid Network for Enhanced Image Registration

Tai Ma (East China Normal University), Ying Wen (East China Normal University)

Image TranslationSegmentationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography

🎯 What it does: A residual pyramid network based on iterative reasoning (IIRP-Net) is proposed for unsupervised medical image deformation registration.

iKUN: Speak to Trackers without Retraining

Yunhao Du (Beijing University of Posts and Telecommunications), Fei Su (Beijing University of Posts and Telecommunications)

Object TrackingVision Language ModelImageVideo

🎯 What it does: A pluggable 'Knowledge Unified Network (iKUN)' is proposed to achieve query-based tracking for any multi-object tracker;

Image Neural Field Diffusion Models

Yinbo Chen (University of California San Diego), Michael Gharbi (Adobe Research)

GenerationData SynthesisDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkImage

🎯 What it does: A diffusion model for image neural fields is proposed, constructing a diffusion framework capable of generating high-quality images at any resolution.

Image Processing GNN: Breaking Rigidity in Super-Resolution

Yuchuan Tian (Peking University), Yunhe Wang (Huawei)

RestorationSuper ResolutionGraph Neural NetworkImage

🎯 What it does: A method for image super-resolution using Graph Neural Networks (IPG) is proposed, breaking the fixed aggregation patterns of traditional convolution and window attention.

Image Restoration by Denoising Diffusion Models with Iteratively Preconditioned Guidance

Tomer Garber (Open University of Israel), Tom Tirer (Bar-Ilan University)

RestorationSuper ResolutionDiffusion modelImage

🎯 What it does: This paper proposes a new iterative pre-adjustment guided method for image restoration tasks, achieving untrained inversion by combining a pre-trained denoiser.

Image Sculpting: Precise Object Editing with 3D Geometry Control

Jiraphon Yenphraphai, Saining Xie

GenerationData SynthesisDiffusion modelImageMeshBenchmark

🎯 What it does: The paper proposes the Image Sculpting framework, which generates a 3D mesh from a single image, supporting precise editing such as pose, rotation, translation, stitching, and sculpting, before re-rendering back to 2D.

Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation

Ji-Jia Wu (National Taiwan University), Yen-Yu Lin (National Yang Ming Chiao Tung University)

SegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality

🎯 What it does: A framework for image-text collaborative decomposition (CoDe) is proposed for text-supervised semantic segmentation; it decomposes images into regions and text into word segments, achieving region-word alignment through contrastive learning without the need for pixel-level annotations.

Image-to-Image Matching via Foundation Models: A New Perspective for Open-Vocabulary Semantic Segmentation

Yuan Wang (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)

SegmentationRetrievalDiffusion modelImage

🎯 What it does: This paper proposes an unsupervised open vocabulary semantic segmentation framework RIM, treating region classification as an image-to-image matching problem.

ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object

Chenshuang Zhang (KAIST), Chengzhi Mao (MILA)

ClassificationData SynthesisConvolutional Neural NetworkTransformerPrompt EngineeringDiffusion modelImageBenchmark

🎯 What it does: Using text-guided diffusion models to synthesize images with diverse backgrounds, textures, and materials, we constructed the ImageNet-D test set to evaluate the robustness of visual models.

Imagine Before Go: Self-Supervised Generative Map for Object Goal Navigation

Sixian Zhang (Chinese Academy of Sciences), Shuqiang Jiang (Chinese Academy of Sciences)

GenerationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodality

🎯 What it does: A self-supervised generative mapping (SGM) framework is proposed, which utilizes local observations and the general knowledge of large language models to complete unobserved semantic map areas, thereby providing more complete environmental information for the Object Navigation (ObjectNav) task.

Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification

Kaijie Ren (Chongqing University), Lei Zhang (Chongqing University)

RecognitionRetrievalKnowledge DistillationConvolutional Neural NetworkContrastive LearningImageMultimodality

🎯 What it does: An Implicit Discriminative Knowledge Learning (IDKL) network is proposed to fully exploit and utilize implicit discriminative information in the visible-infrared cross-modal person re-identification task, enhancing the discriminative power of shared features.

Implicit Event-RGBD Neural SLAM

Delin Qu (Fudan University), Xuelong Li (Fudan University)

Depth EstimationOptimizationRobotic IntelligenceNeural Radiance FieldSimultaneous Localization and MappingImageMultimodality

🎯 What it does: Proposes the EN-SLAM event-RGBD implicit neural SLAM framework, which achieves the fusion of events and RGBD through a unified radiance field and a differentiable Camera Response Function (CRF), enabling robust localization and dense reconstruction.

Implicit Motion Function

Yue Gao (Microsoft Research), Yan Lu (Microsoft Research)

RestorationGenerationCompressionGenerative Adversarial NetworkVideo

🎯 What it does: Proposes the Implicit Motion Function (IMF), which implicitly models the motion between video frames through low-dimensional latent tokens to achieve video reconstruction, compression, and editing.

IMPRINT: Generative Object Compositing by Learning Identity-Preserving Representation

Yizhi Song (Purdue University), Daniel Aliaga (Purdue University)

GenerationData SynthesisDiffusion modelImageVideo

🎯 What it does: This paper proposes a two-stage diffusion model called IMPRINT, which first pre-trains a viewpoint-invariant identity-preserving encoder, and then seamlessly integrates background and geometry based on this.

Improved Baselines with Visual Instruction Tuning

Haotian Liu (University of Wisconsin Madison), Yong Jae Lee (University of Wisconsin Madison)

RecognitionGenerationRetrievalOptimizationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality

🎯 What it does: On the visual instruction tuning framework LLaVA, a more efficient LLaVA-1.5 model was built by improving the visual-language connector, incorporating academic task data, and formatting prompts;

Improved Implicit Neural Representation with Fourier Reparameterized Training

Kexuan Shi (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)

🎯 What it does: An improved implicit neural representation (INR) method is proposed to alleviate the low-frequency bias problem of multilayer perceptrons (MLPs) through Fourier reparameterization training.

Improved Self-Training for Test-Time Adaptation

Jing Ma (Huazhong University of Science and Technology)

ClassificationObject DetectionSegmentationDomain AdaptationImage

🎯 What it does: An improved self-training method called IST is proposed for adapting to the target domain under unsupervised data conditions during testing, achieving a stable adaptive process through lightweight augmentation, pseudo-label correction, and parameter moving average.

Improved Visual Grounding through Self-Consistent Explanations

Ruozhen He (Rice University), Vicente Ordonez (University of California, Irvine)

RecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality

🎯 What it does: In a weakly supervised setting with only image-text pairs, the model is fine-tuned using the SelfEQ loss to enhance visual grounding performance by generating synonymous phrases and ensuring that the model produces the same GradCAM attention maps for the original sentence and its synonymous sentence.

Improved Zero-Shot Classification by Adapting VLMs with Text Descriptions

Oindrila Saha (University of Massachusetts), Subhransu Maji (University of Massachusetts)

ClassificationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageText

🎯 What it does: In this work, the authors enhance the zero-shot classification performance of visual-language models (VLMs) like CLIP by fine-tuning them on a fine-grained domain using fine-grained category descriptions generated by large language models (LLMs) in conjunction with existing large-scale fine-grained image classification data.

Improving Bird's Eye View Semantic Segmentation by Task Decomposition

Tianhao Zhao (Wuhan University), Yutian Lin (Wuhan University)

SegmentationAutonomous DrivingComputational EfficiencyTransformerAuto EncoderImage

🎯 What it does: The paper proposes a two-stage task decomposition method (Task Decomposition, TaDe) for semantic segmentation from monocular images to bird's eye view (BEV). It first uses a BEV autoencoder to learn BEV structural knowledge, then maps RGB images to the BEV latent space for alignment, and finally utilizes a frozen BEV decoder to obtain segmentation results.

Improving Depth Completion via Depth Feature Upsampling

Yufei Wang (Northwestern Polytechnical University), Yuchao Dai

RestorationDepth EstimationAutonomous DrivingConvolutional Neural NetworkImagePoint Cloud

🎯 What it does: This paper proposes a Deep Feature Upsampling Network (DFU) that utilizes dense features within an encoder-decoder network along with a Confidence-Guided Module (CGM) to achieve upsampling of deep features from low resolution to high resolution, significantly improving the accuracy of sparse to dense depth completion.

Improving Distant 3D Object Detection Using 2D Box Supervision

Zetong Yang (Chinese University of Hong Kong), Jose M. Alvarez (NVIDIA)

Object DetectionAutonomous DrivingKnowledge DistillationPoint Cloud

🎯 What it does: In long-range 3D object detection, the authors train the model using only 2D box annotations and propose the LR3D framework, achieving 3D detection of objects at distances greater than 40m.

Improving Generalization via Meta-Learning on Hard Samples

Nishant Jain (Google Research India), Pradeep Shenoy (Google Research India)

ClassificationOptimizationMeta LearningImageBiomedical Data

🎯 What it does: Proposes meta-optimization of the validation set within the Learning Reweighting (LRW) framework, utilizing hard samples from the training set as the validation set to enhance the classifier's generalization performance on both internal and external distributions.

Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names

Yapeng Li (Wuhan University), Bo Du (Wuhan University)

ClassificationRecognitionSupervised Fine-TuningImage

🎯 What it does: A Generalized Zero-Shot Learning (GZSL) framework DSECN is proposed, which enhances the recognition of similar and dissimilar unknown categories by introducing diverse semantics of external category names and aligning them to the visual space.

Improving Graph Contrastive Learning via Adaptive Positive Sampling

Jiaming Zhuo (Hebei University of Technology), Liang Yang (Sun Yat-sen University)

OptimizationRepresentation LearningGraph Neural NetworkContrastive LearningImageGraph

🎯 What it does: An adaptive positive sampling framework HEATS based on self-expression learning is proposed, which achieves global positive sample mining by constructing a block diagonal and unitary positive sample matrix, and alternately optimizes with contrastive learning.

Improving Image Restoration through Removing Degradations in Textual Representations

Jingbo Lin (Harbin Institute of Technology), Wangmeng Zuo (Huawei Cloud Computing Co., Ltd.)

RestorationConvolutional Neural NetworkTransformerDiffusion modelImageText

🎯 What it does: This paper proposes a novel method that first removes image degradation in the text space to generate a clean guiding image, and then uses this guiding image to assist in image restoration.

Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments

Yinhua Piao (Seoul National University), Sun Kim (Seoul National University)

Domain AdaptationDrug DiscoveryGraph Neural NetworkGraph

🎯 What it does: A hierarchical semantic environment generation framework is designed to enhance the OOD generalization performance of graph structures through hierarchical random subgraph extraction and environment inference.

Improving Physics-Augmented Continuum Neural Radiance Field-Based Geometry-Agnostic System Identification with Lagrangian Particle Optimization

Takuhiro Kaneko (NTT Corporation)

OptimizationNeural Radiance FieldVideoPhysics Related

🎯 What it does: This paper proposes the introduction of Lagrangian Particle Optimization (LPO) within the PAC-NeRF framework to enhance geometric-agnostic system identification under sparse viewpoints.

Improving Plasticity in Online Continual Learning via Collaborative Learning

Maorong Wang (University of Tokyo), Toshihiko Yamasaki (University of Tokyo)

ClassificationKnowledge DistillationConvolutional Neural NetworkImage

🎯 What it does: In the framework of Online Continual Learning, a strategy based on Collaborative Learning and Distillation Chain (CCL-DC) is proposed, aiming to significantly enhance the model's plasticity and final accuracy.

Improving Semantic Correspondence with Viewpoint-Guided Spherical Maps

Octave Mariotti (University of Edinburgh), Hakan Bilen (University of Edinburgh)

SegmentationPose EstimationTransformerContrastive LearningImage

🎯 What it does: A semantic correspondence method is proposed that incorporates a weak 3D spherical prior based on self-supervised visual features, along with a new evaluation metric called KAP.

Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment

Muhammad Sohail Danish (Mohamed bin Zayed University of Artificial Intelligence), Mohsen Ali (Information Technology University of Punjab)

Object DetectionDomain AdaptationConvolutional Neural NetworkImage

🎯 What it does: A single-source domain generalization method for object detection is proposed by enhancing visual corrosion on a single source domain and aligning the detection results (classification and localization) for consistency between the enhanced images and the original images.

Improving Spectral Snapshot Reconstruction with Spectral-Spatial Rectification

Jiancheng Zhang (Northwestern Polytechnical University), Yin-Ping Zhao (Northwestern Polytechnical University)

RestorationTransformerImage

🎯 What it does: A Spectral-Spatial Rectification (SSR) method is proposed, enhancing the quality of spectral snapshot reconstruction using Window Self-Attention (WSSA) and Spatial Alignment Blocks (ARB).

Improving Subject-Driven Image Synthesis with Subject-Agnostic Guidance

Kelvin C.K. Chan (Google), Huisheng Wang (Google)

GenerationData SynthesisDiffusion modelImageText

🎯 What it does: A Subject-Agnostic Guidance (SAG) method is proposed to address the issue of models overly focusing on reference images while neglecting text descriptions in subject-driven text-to-image generation.

Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation

Haojie Zhang (South China University of Technology), Kui Jia

SegmentationDomain AdaptationContrastive LearningImageBiomedical Data

🎯 What it does: In response to the insufficient robustness of Segment-Anything (SAM) when facing distribution shifts, a passive data, low memory consumption self-supervised adaptation framework is proposed. This framework utilizes weak supervision cues (bounding boxes, points, rough masks) for self-training and anchor point regularization of SAM, and further enhances performance through contrastive loss and low-rank fine-tuning.

Improving Training Efficiency of Diffusion Models via Multi-Stage Framework and Tailored Multi-Decoder Architecture

Huijie Zhang (University of Michigan), Qing Qu (University of Michigan)

GenerationComputational EfficiencyDiffusion modelImage

🎯 What it does: A multi-stage framework and multi-decoder U-Net structure are proposed to improve the training and sampling efficiency of diffusion models.

Improving Transferable Targeted Adversarial Attacks with Model Self-Enhancement

Han Wu (Sun Yat-sen University), Zibin Zheng (Sun Yat-sen University)

Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This paper proposes a self-improvement method for source models aimed at enhancing the transferability of target attacks.

Improving Unsupervised Hierarchical Representation with Reinforcement Learning

Ruyi An (Nanyang Technological University), Mingyuan Zhou (University of Texas at Austin)

Representation LearningReinforcement LearningAuto EncoderImage

🎯 What it does: A hierarchical variational autoencoder (VAE) training framework based on reinforcement learning is proposed, aimed at eliminating posterior collapse and learning interpretable hierarchical representations.

Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping

Hyeongjun Kwon (Yonsei University), Kwanghoon Sohn (Korea Institute of Science and Technology)

ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerContrastive LearningImage

🎯 What it does: The Hi-Mapper module is proposed, which constructs a visual hierarchy through a probabilistic tree and maps the hierarchy to hyperbolic space, ultimately enhancing the structured representation capability of pre-trained DNNs.

In Search of a Data Transformation That Accelerates Neural Field Training

Junwon Seo (Pohang University of Science and Technology), Jaeho Lee (Pohang University of Science and Technology)

Super ResolutionOptimizationComputational EfficiencyNeural Radiance FieldImage

🎯 What it does: This study explores the possibility of transforming data to accelerate neural field training and finds that random pixel permutation can significantly improve training speed.

In-Context Matting

He Guo (Huazhong University of Science and Technology), Hao Lu (Huazhong University of Science and Technology)

SegmentationGenerationTransformerDiffusion modelImage

🎯 What it does: The task of 'in-context matting' is proposed, which automatically generates alpha masks for a set of target images of the same category using a single reference image and its point/brush/mask prompts, and develops the IconMatting model based on Stable Diffusion;

In-distribution Public Data Synthesis with Diffusion Models for Differentially Private Image Classification

Jinseong Park (Seoul National University), Jaewook Lee (Seoul National University)

ClassificationGenerationData SynthesisSafty and PrivacyDiffusion modelImage

🎯 What it does: This paper proposes the use of diffusion models to synthesize in-distribution (ID) public data, which is then used as warm-up data for differential privacy (DP) image classification, significantly improving the accuracy of private training.

In-N-Out: Faithful 3D GAN Inversion with Volumetric Decomposition for Face Editing

Yiran Xu (University of Maryland), Jia-Bin Huang (University of Maryland)

GenerationData SynthesisNeural Radiance FieldGenerative Adversarial NetworkImageVideo

🎯 What it does: By splitting the input image into two parts: in-distribution (natural faces) and out-of-distribution (occlusions or heavy makeup, etc.), we represent them using three-plane NeRF and achieve 3D GAN inversion and editing through volumetric composite rendering.

In2SET: Intra-Inter Similarity Exploiting Transformer for Dual-Camera Compressive Hyperspectral Imaging

Xin Wang (Beijing Institute of Technology), Hua Huang (Beijing Normal University)

RestorationTransformerImage

🎯 What it does: This paper proposes In2SET, a Transformer-based DCCHI hyperspectral image reconstruction network that utilizes the internal and cross-modal similarities of panchromatic images to enhance reconstruction quality.

InceptionNeXt: When Inception Meets ConvNeXt

Weihao Yu (National University of Singapore), Xinchao Wang (National University of Singapore)

ClassificationSegmentationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage

🎯 What it does: A new CNN architecture called InceptionNeXt is proposed, which utilizes a four-branch Inception depthwise separable convolution (small square convolution, horizontal/vertical strip convolution, and identity mapping) to replace traditional large-kernel depthwise separable convolution, significantly improving training/inference throughput while maintaining or enhancing accuracy.

Incorporating Geo-Diverse Knowledge into Prompting for Increased Geographical Robustness in Object Recognition

Kyle Buettner (University of Pittsburgh), Adriana Kovashka (University of Pittsburgh)

RecognitionObject DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageText

🎯 What it does: This paper studies how to introduce geographically diverse knowledge into object recognition tasks to enhance the model's robustness in different geographical environments, and parameterizes the fine-tuning of CLIP through soft prompting, combining geographical descriptions generated by large language models for knowledge regularization.

Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation

Huyong Wang (Shenzhen University), Jing Qin (The Hong Kong Polytechnic University)

SegmentationKnowledge DistillationImageBiomedical Data

🎯 What it does: Incremental nucleus segmentation in clinical pathological images addresses the problem of catastrophic forgetting without saving old class samples.

Incremental Residual Concept Bottleneck Models

Chenming Shang (Tsinghua University), Yuwang Wang (Tsinghua University)

ClassificationExplainability and InterpretabilityContrastive LearningImage

🎯 What it does: This paper proposes the Incremental Residual Concept Bottleneck Model (Res-CBM), which optimizes vector filling for missing concepts in the base concept library through random initialization, and further transforms these residual vectors into interpretable candidate concepts, thereby enhancing the conceptual integrity and classification performance of CBM.

Infer from What You Have Seen Before: Temporally-dependent Classifier for Semi-supervised Video Segmentation

Jiafan Zhuang (Shantou University), Zhun Fan (Shantou University)

SegmentationAutonomous DrivingRecurrent Neural NetworkVideo

🎯 What it does: A temporal dependency classifier (TDC) is proposed, which utilizes the semantic context of historical frames to enhance classification performance in semi-supervised learning for video semantic segmentation.

Infinigen Indoors: Photorealistic Indoor Scenes using Procedural Generation

Alexander Raistrick (Princeton University), Jia Deng (Princeton University)

GenerationData SynthesisOptimizationImage

🎯 What it does: This paper presents Infinigen Indoors, a comprehensive system for procedurally generating realistic indoor scenes based on Blender.

InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning

Yan-Shuo Liang (Nanjing University), Wu-Jun Li (Nanjing University)

ClassificationRecognitionTransformerSupervised Fine-TuningImage

🎯 What it does: A new parameter-efficient fine-tuning method called InfLoRA is proposed, which utilizes low-rank adaptation to insert parameters and construct subspaces, aiming to eliminate interference from old tasks on new tasks and achieve continual learning.

Infrared Adversarial Car Stickers

Xiaopei Zhu (Tsinghua University), Xiaolin Hu (Tsinghua University)

Object DetectionAutonomous DrivingAdversarial AttackMesh

🎯 What it does: Designed and manufactured 3D full-view car stickers as physical adversarial examples for infrared target detection.

Infrared Small Target Detection with Scale and Location Sensitivity

Qiankun Liu (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)

Object DetectionConvolutional Neural NetworkImage

🎯 What it does: A scale and position-sensitive loss function SLS is proposed, combined with a simplified U-Net model with multi-scale heads, MSHNet, to improve the accuracy of infrared small target detection.

Initialization Matters for Adversarial Transfer Learning

Andong Hua (University of California), Yao Qin (Google)

Domain AdaptationAdversarial AttackTransformerSupervised Fine-TuningImage

🎯 What it does: This study investigates the adversarial robustness in transfer learning and proposes a method to enhance the robustness of downstream tasks through Adversarial Linear Initialization (RoLI).

InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization

Xiefan Guo (Beihang University), Di Huang (Beihang University)

GenerationOptimizationTransformerDiffusion modelImageText

🎯 What it does: The Initial Noise Optimization (INITNO) method is proposed, which enhances the performance of text-to-image diffusion models in adhering to prompts by optimizing noise in the initial latent space.

Ink Dot-Oriented Differentiable Optimization for Neural Image Halftoning

Hao Jiang (Peking University), Yadong Mu (Peking University)

Image TranslationOptimizationConvolutional Neural NetworkImage

🎯 What it does: A neural image halftoning method that can directly control the position of ink dots is proposed, allowing for fine image reproduction through the adjustment of ink dots.

Inlier Confidence Calibration for Point Cloud Registration

Yongzhe Yuan (Xidian University), Wenping Ma (Xidian University)

OptimizationGraph Neural NetworkPoint Cloud

🎯 What it does: A method is proposed to calibrate inlier confidence (ICC) in point cloud registration through geometric constraints, achieving end-to-end training in an unsupervised framework.

InNeRF360: Text-Guided 3D-Consistent Object Inpainting on 360-degree Neural Radiance Fields

Dongqing Wang (École Polytechnique Fédérale de Lausanne), Sabine Süsstrunk (École Polytechnique Fédérale de Lausanne)

RestorationSegmentationGenerationDiffusion modelNeural Radiance FieldImage

🎯 What it does: This paper presents InNeRF360, a text instruction-based 360° NeRF scene object removal and consistency filling system that can automatically generate float-free, visually consistent 3D filling areas.

Insect-Foundation: A Foundation Model and Large-scale 1M Dataset for Visual Insect Understanding

Hoang-Quan Nguyen (University of Arkansas), Khoa Luu (University of Arkansas)

ClassificationObject DetectionTransformerContrastive LearningImageMultimodality

🎯 What it does: A 1M-scale insect image dataset (Insect-1M) and an insect foundation model (Insect-Foundation) based on Patch-wise Relevant Attention are proposed, supporting insect vision, detection, and vision-language tasks.

Insights from the Use of Previously Unseen Neural Architecture Search Datasets

Rob Geada (Newcastle University), A. Stephen McGough (Newcastle University)

Neural Architecture SearchConvolutional Neural NetworkTabular

🎯 What it does: This paper designs and releases eight previously unseen datasets and evaluates various NAS methods on them to explore the generalization ability of NAS on unknown tasks.

InstaGen: Enhancing Object Detection by Training on Synthetic Dataset

Chengjian Feng (Meituan Inc), Lin Ma (Meituan Inc)

Object DetectionData SynthesisSupervised Fine-TuningDiffusion modelImage

🎯 What it does: By fine-tuning the diffusion model and adding an instance localization head, high-quality synthetic images with instance boxes are generated to train and enhance the performance of object detectors.

Instance Tracking in 3D Scenes from Egocentric Videos

Yunhan Zhao (University of California Irvine), Charless Fowlkes (University of California Irvine)

Object TrackingSimultaneous Localization and MappingVideoMultimodalityBenchmark

🎯 What it does: The IT3DEgo benchmark task and dataset are proposed, focusing on tracking target instances from a first-person perspective (Egocentric) camera in real indoor 3D scenes.

Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

Xiao Lin (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)

Pose EstimationTransformerPoint Cloud

🎯 What it does: This paper proposes an AG-Pose method that learns sparse keypoints and establishes correspondences at the keypoint level for 6D object pose estimation of unseen instances through instance-adaptive keypoint detection and geometric-aware feature aggregation.