Shadow removal under diverse lighting requires disentangling illumination from intrinsic reflectance— a challenge when physical priors are misaligned.
We propose PhaSR with dual-level prior alignment: (1) PAN performs parameter-free illumination correction via Gray-world normalization and log-domain Retinex decomposition, suppressing chromatic bias. (2) GSRA extends differential attention to harmonize depth-derived geometry with DINO-v2 semantics, resolving modal conflicts across illumination conditions.
Experiments demonstrate competitive performance with lower complexity, generalizing to ambient lighting where traditional methods fail.
A multi-scale Transformer encoder-decoder integrates frozen DINO-v2 semantic features and DepthAnything-v2 geometric priors via GSRA's cross-modal differential attention (Arect = Asem - λ·Ageo).
PhaSR: Generalized Image Shadow Removal with Physically Aligned Priors
@article{lee2025phasr,
title={PhaSR: Generalized Image Shadow Removal with Physically Aligned Priors},
author={Lee, Chia-Ming and Lin, Yu-Fan and Hsiao, Yu-Jou and Jung, Jing-Hui and Liu, Yu-Lun and Hsu, Chih-Chung},
journal={arXiv preprint arXiv:XXXX.XXXXX},
year={2025}
}