arXivSub Start free trial

CVPR 2024 Papers — Page 28

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

You'll Never Walk Alone: A Sketch and Text Duet for Fine-Grained Image Retrieval

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

RetrievalGenerative Adversarial NetworkContrastive LearningImageText

🎯 What it does: A framework for unsupervised sketch + text combination retrieval and generation based on CLIP is proposed, which maps sketches to pseudo-word tokens and achieves fine-grained retrieval through sketch-text differentiation and neutral text regularization.

Your Image is My Video: Reshaping the Receptive Field via Image-To-Video Differentiable AutoAugmentation and Fusion

Sofia Casarin (Free University of Bozen-Bolzano), Oswald Lanz (Free University of Bozen-Bolzano)

ClassificationSegmentationNeural Architecture SearchConvolutional Neural NetworkImageVideo

🎯 What it does: A sequence of image transformation generated through differentiable incremental search forms a short video, which reshapes the receptive field of convolutional networks through a lightweight video network, thereby enhancing classification and semantic segmentation performance.

Your Student is Better Than Expected: Adaptive Teacher-Student Collaboration for Text-Conditional Diffusion Models

Nikita Starodubcev (Yandex Research), Artem Babenko (Yandex Research)

GenerationOptimizationKnowledge DistillationDiffusion modelImage

🎯 What it does: An adaptive teacher-student collaboration method is proposed, where a few-step student model is first used to generate images, and then a threshold is applied to decide whether to use the teacher model for further optimization.

Your Transferability Barrier is Fragile: Free-Lunch for Transferring the Non-Transferable Learning

Ziming Hong (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)

Domain AdaptationKnowledge DistillationImage

🎯 What it does: This paper studies the robustness of transferability barriers in Non-Transferable Learning (NTL) models, finding that they generally degrade in third-party domains. It proposes the TransNTL attack method, which can restore target domain performance using only a small amount of source domain data, and presents a defense scheme that incorporates perturbation distillation during the training phase.

Z*: Zero-shot Style Transfer via Attention Reweighting

Yingying Deng (Institute of Automation, Chinese Academy of Sciences), Weiming Dong (Institute of Automation, Chinese Academy of Sciences)

Image TranslationGenerationDiffusion modelImage

🎯 What it does: This paper proposes a zero-shot image style transfer method called Z-STAR, which utilizes a pre-trained diffusion model to generate content and style latent codes through a dual-path reverse process, and achieves untrained style transfer using cross-attention reweighting during the reverse diffusion process.

ZePT: Zero-Shot Pan-Tumor Segmentation via Query-Disentangling and Self-Prompting

Yankai Jiang (Shanghai AI Laboratory), Shaoting Zhang (SenseTime Research)

SegmentationTransformerPrompt EngineeringImageBiomedical DataComputed Tomography

🎯 What it does: The ZePT framework is proposed, which can perform zero-shot tumor segmentation using only annotated organ data, and simultaneously segment known organs and unknown tumors during inference.

ZERO-IG: Zero-Shot Illumination-Guided Joint Denoising and Adaptive Enhancement for Low-Light Images

Yiqi Shi (Harbin Engineering University), Xiaojing Fu (Harbin Engineering University)

RestorationImage

🎯 What it does: This paper proposes a zero-shot, illumination-guided joint denoising and adaptive enhancement method that can process real low-light images without relying on training data and noise distribution.

Zero-Painter: Training-Free Layout Control for Text-to-Image Synthesis

Marianna Ohanyan (Picsart AI Research), Humphrey Shi (Picsart AI Research)

GenerationData SynthesisPrompt EngineeringDiffusion modelImage

🎯 What it does: A training-free, layout-based text-to-image generation framework called Zero-Painter is proposed, which can generate images that conform to the shape and text attributes from object masks, corresponding text descriptions, and global prompts.

Zero-Reference Low-Light Enhancement via Physical Quadruple Priors

Wenjing Wang (Peking University), Jiaying Liu (Peking University)

RestorationGenerationKnowledge DistillationTransformerDiffusion modelImage

🎯 What it does: A zero-reference low-light enhancement framework is proposed, utilizing an optical quadruple prior based on the Kubelka-Munk theory and a pre-trained diffusion model to achieve low-light image enhancement without low-light data training.

Zero-shot Referring Expression Comprehension via Structural Similarity Between Images and Captions

Zeyu Han (Northeastern University), Huaizu Jiang (Northeastern University)

RecognitionObject DetectionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality

🎯 What it does: Proposes a zero-shot representation understanding model that utilizes ChatGPT to parse text triples, CLIP/FLAVA to compute structural similarity of triples, and propagates this to instance-level for visual localization.

Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping

Ruoxi Zhu (Fudan University), Yibo Fan (Fudan University)

Image TranslationRestorationDiffusion modelImage

🎯 What it does: This paper proposes a zero-shot structure-preserving diffusion model that guides the tone mapping from HDR to LDR using structural information.

Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers

Hongjie Wang (Princeton University), Niraj K. Jha (Princeton University)

ClassificationComputational EfficiencyHyperparameter SearchTransformerImage

🎯 What it does: A zero-shot token pruning method called Zero-TPrune is proposed, which can identify and reduce redundant or unimportant tokens by focusing on the attention maps of a pre-trained Transformer without any fine-tuning, significantly reducing inference FLOPs and latency.

ZeroNVS: Zero-Shot 360-Degree View Synthesis from a Single Image

Kyle Sargent (Stanford University), Jiajun Wu (Stanford University)

GenerationData SynthesisDiffusion modelScore-based ModelImage

🎯 What it does: Trained a 3D-aware diffusion model ZeroNVS, achieving 360° view synthesis from a single image, supporting real multi-object scenes.

ZeroRF: Fast Sparse View 360deg Reconstruction with Zero Pretraining

Ruoxi Shi (University of California San Diego), Hao Su (University of California San Diego)

RestorationGenerationData SynthesisConvolutional Neural NetworkNeural Radiance FieldImage

🎯 What it does: We propose ZeroRF, a method for rapid 360° scene reconstruction without pre-training and capable of operating under sparse viewpoints.

ZeroShape: Regression-based Zero-shot Shape Reconstruction

Zixuan Huang (University of Illinois at Urbana-Champaign), James M. Rehg (University of Illinois at Urbana-Champaign)

GenerationDepth EstimationTransformerAuto EncoderImageMeshBenchmark

🎯 What it does: This paper proposes ZeroShape, a regression-based zero-shot 3D shape reconstruction method, and establishes a unified large-scale real evaluation benchmark.

ZONE: Zero-Shot Instruction-Guided Local Editing

Shanglin Li (Beihang University), Baochang Zhang (Beihang University)

Image TranslationGenerationDiffusion modelImage

🎯 What it does: A zero-shot instruction-driven local image editing method is proposed, which uses natural language instructions to locate and edit image regions while keeping non-edited areas unchanged. It supports operations such as addition, deletion, modification, and multi-round editing.