AI Paper Atlas

Research map / Image generation / Image restoration

Leveraging Inpainting for Single-Image Shadow Removal

Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, Song Wang

ICCV 2023 · Image generation › Image restoration · approach: Dehazing and deraining · 39 citations (OpenAlex)

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 changedWhereMeasuredEffect
add inpainting pretraining before shadow removal fine-tuning (ablation)naive encoder-decoder · ISTD+RMSE (all): 4.216 -> 3.527Improved
add adaptive fusion block (ablation)two-encoder shadow removal network · ISTD+RMSE (all): 3.627 -> 3.398Improved
use two separate encoders instead of concatenating shadow and masked inputs (ablation)adaptive fusion network · ISTD+RMSE (all): 3.581 -> 3.398Improved
combine shadow removal with inpainting rather than de-shadowing by inpainting alone (ablation)adaptive fusion network · ISTD+RMSE (all): 5.992 -> 3.398Improved
add sigmoid to predicted fusion weights (ablation)adaptive fusion block · ISTD+RMSE (all): 3.597 -> 3.398Improved
keep both weights W1 and W2 in fusion (ablation)adaptive fusion block · ISTD+RMSE (all): 3.432 -> 3.398 (w/o W1); 3.473 -> 3.398 (w/o W2)Improved
pretrain on larger inpainting dataset (full Places2 vs 10^4 images) (ablation)adaptive fusion network · ISTD+RMSE (all): 3.476 -> 3.398Improved
pretrain on any inpainting dataset (CelebA, Paris-Street-View, Places2 10^4) vs none (ablation)adaptive fusion network · ISTD+RMSE (all): 3.860 -> 3.557 / 3.511 / 3.476Improved
inpainting-pretrain, then fine-tune on only 10% of ISTD+naive encoder-decoder · ISTD+RMSE shadow-free / shadow region: 3.07 vs 3.80 (AEFNet); 7.67 vs 7.9 (SP+M-Net)Improved
use longer inpainting pretraining (stronger inpainting model) (ablation)naive encoder-decoder · ISTD+RMSEImproved
add inpainting pretraining (ablation)naive encoder-decoder · ISTD+generalization (RMSE gap, IIW)Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy