STICT for Video Shadow Detection

Code and Dataset for our CVPR 2022 paper "Video Shadow Detection via Spatio-Temporal Interpolation Consistency Training"

VIdeo ShAdow Detection dataset (VISAD)

VISAD is consisted of 82 videos and was divided into two parts according to semantic of scenes: the Driving Scenes (VISAD-DS) part and the Moving Object Scenes (VISAD-MOS) part, denoted as DS and MOS respectively.

It is available at Google Drive.

scenes	videos/annotated	frames/annotated	resolution
DS-all	47 / 17	7953 / 2881	1280×720
DS-test	13 / 13	2190 / 2190	1280×720
MOS-all	34 / 16	4613 / 1307	(530-1920)×(360-1080)
MOS-test	13 / 13	873 / 873	1920×1080,1600×900

evaluation over predictions

Run python evaluate.py

important arguments:
-gp, --gt_path ground truth path
-pp, --pred_path your predicitons path

Our pretrained shadow maps is available here(DS, MOS, ViSha)

Spatio-Temporal Interpolation Consistency Training

Requirement

cuda (10.0)
Python (3.6)
PyTorch (1.1.0)
spatial-correlation-sampler (0.0.8)
Flownet (2.0)

Download dataset

Download the following datasets and unzip them into ./data folder

SBU (it can refer MTMT)
DS
MOS

Testing

Our pretrained model is available here

Run python test.py

important arguments:
--trained_model trained model path (default:'./DS')
--dataset_path your test set path (default: './data/DS/test/')
--dataset_txt_path your test set list path (default: './data/DS/test/test.txt')

Training

Download pretrained models (ResNet and FlowNet) into ./pretrained_model folder
Run python train.py

important arguments:
--target_domain (options: 'DS_U', 'MOS_U', 'ViSha') (default: 'DS_U')
--dataset_U_path your video domain dataset path (default: './data/DS/train/')

yihong-97/STICT