stcgan

This is an unofficial implementation of CVPR 2018 "Stacked Conditional Generative Adversarial Networks for Jointly Learning Shadow Detection and Shadow Removal"

Q&A

1. what's the difference between natural scene and documents shadow removal?

EXPERIMENT

1. stcgan_fusion.py

   ---

   | S_img | -----> | STCGAN | -----> | N1_img | -----> | | --------- ---------- ---------- | | --------- | | Blending | -----> | N_img | | ---------- ---------- | | --------- ------------> | ACCV | -----> | N2_img | -----> | | ---------- ---------- ------------

2. pix2pix.py

   ---

   | S_img | -----> | ACCV | -----> | N1_img | -----> | STCGAN | -----> | N_img |

   ---

3. stcgan_concat.py

   ---

   | S_img | -----> | ACCV | -----> | N1_img | -----> | | --------- ---------- ---------- | | --------- | | STCGAN | -----> | N_img | | | | --------- ----------------------------------------------------> | | ----------

TODO

1. find the best model
2. record the iteration and save it in to file, so it can be resumed when reloaded
3. plot the training loss

Execution command

1. train model ./run.sh --mode train --gpu_id 0 --epochs 500 --root_dir /media/yslin/SSD_DATA/research/stcgan/ --batch_size 8 --dataset_name DSRD_BW ./run.sh --mode train --gpu_id 0 --epochs 500 --root_dir /media/yslin/SSD_DATA/research/stcgan/ --batch_size 8 --dataset_name DSRD_all

2. test model ./run.sh --mode test --load_model --gpu_id 0 --root_dir /media/yslin/SSD_DATA/research/stcgan/ --batch_size_test 10 --dataset_name DSRD_aligned --model_name latest ./run.sh --mode test --load_model --gpu_id 0 --root_dir /media/yslin/SSD_DATA/research/stcgan/ --batch_size_test 10 --dataset_name DSRD_aligned

3. evaluate all model python src/eval_all_model.py --root_dir /media/yslin/SSD_DATA/research/stcgan/ --batch_size_test 10 --lrG 0.001 --lrD 0.001 --dataset_name DSRD_aligned python src/eval_all_model.py --root_dir /media/yslin/SSD_DATA/research/stcgan/ --batch_size_test 10 --lrG 0.001 --lrD 0.001 --dataset_name ISTD

4. eval result python src/eval.py ISTD /media/yslin/SSD_DATA/research/stcgan/processed_dataset/ISTD/result/Guo/ python src/eval.py ISTD /media/yslin/SSD_DATA/research/stcgan/processed_dataset/ISTD/result/Yang/ python src/eval.py ISTD /media/yslin/SSD_DATA/research/stcgan/processed_dataset/ISTD/result/Gong/ python src/eval.py ISTD /media/yslin/SSD_DATA/research/stcgan/processed_dataset/ISTD/result/ST-CGAN/ python src/eval.py ISTD /media/yslin/SSD_DATA/research/stcgan/task/stcgan_lrG_0.001_lrD_0.001/ISTD/result/non_shadow

   python src/eval.py DSRD_aligned /media/yslin/SSD_DATA/research/processed_dataset/DSRD_aligned/test/shadow python src/eval.py DSRD_aligned /media/yslin/SSD_DATA/research/processed_dataset/DSRD_aligned/result/ACCV2016 python src/eval.py DSRD_all /media/yslin/SSD_DATA/research/stcgan/task/stcgan_lrG_0.001_lrD_0.001/DSRD_all/result/non_shadow

   python src/eval.py Blender39_aligned /media/yslin/SSD_DATA/research/processed_dataset/Blender39_aligned/test/shadow

5. visualize grid result python src/displayPairDSRD.py