Research map / Detection and segmentation / Open-set and few-shot detection
Perceptual Artifacts Localization for Inpainting
Lingzhi Zhang, Yuqian Zhou, Connelly Barnes, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, Jianbo Shi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the perfectly filled images (no artifact labels) to training (ablation) | PSPNet segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 43.83 -> 46.04 | Improved |
| start from ADE20K pretrained weights (ablation) | PSPNet segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 44.93 -> 46.04 | Improved |
| concatenate the hole mask to the input (ablation) | PSPNet segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 46.04 -> 45.96 | No clear effect |
| add pseudo-label pretraining on 100K unlabeled images (ablation) | PSPNet segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 46.04 -> 46.44 | Improved |
| add real images with empty masks to pseudo pretraining (ablation) | PSPNet segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 46.44 -> 46.77 | Improved |
| use Swin-B backbone with UperNet head instead of ResNet-50 and PSPNet (ablation) | segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 46.04 -> 44.20 | Hurt |
| use HRNet head instead of PSPNet head (ablation) | segmenter · Perceptual artifacts segmentation (test, 479 images) | IoU: 46.04 -> 41.35 | Hurt |
| use PAR instead of PSNR to rank inpainting methods vs human preference | PAR metric · Places2 object removal pairs (1966 strongly preferred pairs) | agreement with human preference: 41.50% -> 72.89% | Improved |
| use PAR instead of MUSIQ to rank inpainting methods vs human preference | PAR metric · Places2 object removal pairs (1966 strongly preferred pairs) | agreement with human preference: 61.28% -> 72.89% | Improved |
| use PAR instead of LPIPS | PAR metric · Places2 object removal pairs (1966 strongly preferred pairs) | agreement with human preference: 38.55% -> 72.89% | Improved |
| add iterative fill on predicted artifact regions (5th refill) to EdgeConnect | EdgeConnect · User study (500 images) | user preference: 10.6% preferred original -> 37.8% preferred iterative | Improved |
| add iterative fill on predicted artifact regions (5th refill) to LaMa | LaMa · User study (500 images) | user preference: 1.8% preferred original -> 30.0% preferred iterative | Improved |
Closest papers by meaning
- Perceptual Artifacts Localization for Image Synthesis Tasks (ICCV 2023 · similarity 0.91)
- High-Resolution Image Inpainting with Iterative Confidence Feedback and Guided Upsampling (ECCV 2020 · similarity 0.89)
- Inst-Inpaint: Instructing to Remove Objects with Diffusion Models (IJCV 2026 · similarity 0.87)
- InpDiffusion: Image Inpainting Localization via Conditional Diffusion Models (AAAI 2025 · similarity 0.87)
- Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency (CVPR 2025 · similarity 0.86)
- Can We Model the Artifacts Explicitly? Disentangle Artifacts via Pairwise Edit Relations for Image Manipulation Localization (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
