Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, Chi-Wing Fu

Instance Shadow Detection aims to find shadow instances, object instances and shadow-object associations; this task benefits many vision applications, such as light direction estimation and photo editing.

To approach this task, we first compile a new dataset with the masks for shadow instances, object instances, and shadow-object associations. We then design an evaluation metric for quantitative evaluation of the performance of instance shadow detection. Further, we design a single-stage detector to perform instance shadow detection in an end-to-end manner, where the bidirectional relation learning module and the deformable maskIoU head are proposed in the detector to directly learn the relation between shadow instances and object instances and to improve the accuracy of the predicted masks.

[📄 TPAMI - SSISv2] [📄 CVPR - SSIS] [👇🏼 Video]

pip install -r requirement.txt

Note that we tested on CUDA10.2 / PyTorch 1.6.0, CUDA11.1-11.5 / PyTorch 1.8.0 - 1.11.0 and Colab.

This repo is implemented on AdelaiDet, so first build it with:

$ cd SSIS
$ python setup.py build develop

Download the dataset SOBA_v2.zip, model_ssis_final.pth, and model_ssisv2_final.pth from Google drive. Put dataset file in the ../dataset/ and put pretrained model in the tools/output/SSIS_MS_R_101_bifpn_with_offset_class/ or tools/output/SSISv2_MS_R_101_bifpn_with_offset_class_maskiouv2_da_bl/.

To evaluate the results, try the command example:

$ cd demo
$ python demo.py --input ./samples

# SSIS
$ cd tools
$ python train_net.py \
    --config-file ../configs/SSIS/MS_R_101_BiFPN_with_offset_class.yaml \
    --num-gpus 2 

# SSISv2 requires more GPU memory. We trained it on RTX 3090
$ cd tools
$ python train_net.py \
    --config-file ../configs/SSIS/MS_R_101_BiFPN_SSISv2.yaml \
    --num-gpus 1

# SSIS
$ python train_net.py \
    --config-file ../configs/SSIS/MS_R_101_BiFPN_with_offset_class.yaml \
    --num-gpus 2 --resume --eval-only
$ python SOAP.py --path PATH_TO_YOUR_DATASET/SOBA \ 
    --input-name ./output/SSIS_MS_R_101_bifpn_with_offset_class
# SSISv2
$ python train_net.py \
    --config-file ../configs/SSIS/MS_R_101_BiFPN_SSISv2.yaml \
    --num-gpus 1 --resume --eval-only
$ python SOAP.py --path PATH_TO_YOUR_DATASET/SOBA \ 
    --input-name ./output/SSISv2_MS_R_101_bifpn_with_offset_class_maskiouv2_da_bl

If you use LISA, SSIS, SSISv2, SOBA, or SOAP, please use the following BibTeX entry.

@ARTICLE{Wang_2022_TPAMI,  
author    = {Wang, Tianyu and Hu, Xiaowei and Heng, Pheng-Ann and Fu, Chi-Wing}, 
journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence},   
title     = {Instance Shadow Detection with A Single-Stage Detector},   
year      = {2022},  
volume    = {},  
number    = {},  
pages     = {1-14},  
doi       = {10.1109/TPAMI.2022.3185628}
}

@InProceedings{Wang_2021_CVPR,
author    = {Wang, Tianyu and Hu, Xiaowei and Fu, Chi-Wing and Heng, Pheng-Ann},
title     = {Single-Stage Instance Shadow Detection With Bidirectional Relation Learning},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month     = {June},
Year      = {2021},
pages     = {1-11}
}

@InProceedings{Wang_2020_CVPR,
author    = {Wang, Tianyu and Hu, Xiaowei and Wang, Qiong and Heng, Pheng-Ann and Fu, Chi-Wing},
title     = {Instance Shadow Detection},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month     = {June},
year      = {2020}
}