Code for our ICCV 2021 paper "Mitigating Intensity Bias in Shadow Detection via Feature Decomposition and Reweighting"
conda env create -f env.yaml
To use CRF refinement, you will need to mannually install pydensecrf.
WARNING: To reproduce the results reported in our paper, please make sure major pacakges (pytorch, opencv, etc.) are with the same version speficified in env.yaml
.
- download the checkpoint from here.
- specify data_root, and run
python test.py
. - run
python crf_refine.py
. - check the results w/ and w/o CRF refinement in
test/raw
andtest/crf
respectively
BER scores are specified below.
SBU | UCF | ISTD | |
---|---|---|---|
w/o CRF | 3.27 | 7.42 | 1.53 |
w/ CRF | 3.04 | 7.28 | 1.55 |
You can access qualitative reseults from BaiduNetDisk (passcode:4j3i).
- move the logic of brightness shift to the dataset class; rewrite dataset class.
- remove feature extraction hook, use segmentation_models.pytorch encoder instead.
- use timm's register_model decorator