Research map / Image generation / Autoregressive and latent generation
Large Occluded Human Image Completion via Image-Prior Cooperating
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LOHC, a two-stage network that jointly completes a human image and its segmentation map (image-prior cooperative completion), using a transformer coarse network and an FFC-based refinement network with a Dynamic Fusion Module.
Problem
Completing large occluded human body images is hard because human shape variation and human-viewer sensitivity to artifacts make prior-then-image completion pipelines fragile when parsing maps are inaccurate.
Approach
Instead of completing the segmentation prior first and then the image, complete the image and the human segmentation map simultaneously in one unified model, with a transformer MAE-style coarse network and a DFM that selectively fuses coarse and input features at channel and spatial levels.
Key results
- On the Center mask (75.00%), LOHC reaches 14.931 whole-image PSNR and 23.722 human-area PSNR, the best among compared methods.
- In a user study of 10 scorers over 50 images, LOHC was preferred most often on every mask set (e.g., 82 votes on Center masks vs 11 for LaMa).
- Ablation shows removing the coarse network drops PSNR from 19.88 to 17.92 and removing the DFM drops it to 17.20 on VOC masks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the human segmentation prior (ablation) | LOHC · VOC masks | PSNR: 18.9394 -> 19.8768 | Helped |
| Use cooperative completion to supervise and complete the segmentation prior (ablation) | LOHC · VOC masks | PSNR: 19.1940 -> 19.8768 | Helped |
| Add the coarse network and its channel-attention fusion (ablation) | LOHC · VOC masks | PSNR: 17.9160 -> 19.8768 | Helped |
| Use separate coarse-network encoders for the image and human area (ablation) | LOHC · VOC masks | PSNR: 18.1886 -> 19.8768 | Helped |
| Use separate encoders rather than jointly encoding the image and human area (ablation) | LOHC · VOC masks | PSNR: 19.0473 -> 19.8768 | Helped |
| Encode the human area separately instead of using both encoders for the image only (ablation) | LOHC · VOC masks | PSNR: 19.5710 -> 19.8768 | Helped |
| Add the dynamic fusion module (ablation) | LOHC · VOC masks | PSNR: 17.2023 -> 19.8768 | Helped |
| Add channel attention to the dynamic fusion module (ablation) | LOHC · VOC masks | PSNR: 19.3467 -> 19.8768 | Helped |
| Add spatial attention to the dynamic fusion module (ablation) | LOHC · VOC masks | PSNR: 19.7791 -> 19.8768 | Helped |
| Add global discriminators (ablation) | LOHC · VOC masks | PSNR: 19.7978 -> 19.8768 | Helped |
Limitations
- Evaluated only on the AHP dataset for training and on a limited set of mask types
- Relies on a U-Net to obtain human segmentation maps, and the approach is not tested with diffusion backbones
Experimental setup
Datasets: AHP
Benchmarks: AHP validation set VOC masks XPIE masks HKU-IS masks
Metrics: PSNR SSIM LPIPS FID user preference
Baselines: LaMa Co-Mod-GAN TFill CR-Fill PBHC
Builds on: Masked Autoencoder LaMa PatchGAN TFill
Closest papers by meaning
- Prior Based Human Completion (CVPR 2021 · similarity 0.89)
- RoGSplat: Learning Robust Generalizable Human Gaussian Splatting from Sparse Multi-View Images (CVPR 2025 · similarity 0.84)
- DPMesh: Exploiting Diffusion Prior for Occluded Human Mesh Recovery (CVPR 2024 · similarity 0.84)
- Dynamic Feature Pruning and Consolidation for Occluded Person Re-identification (AAAI 2024 · similarity 0.84)
- Using Diffusion Priors for Video Amodal Segmentation (CVPR 2025 · similarity 0.84)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.84)
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.