Assume all data is under directory $DATA_ROOT
Download videos and images with the following script:
bash download_video.sh youtube_ids.txt $DATA_ROOT
,where preprocess/youtube_ids.txt contains all youtube video ids used to train PedGen.
Note that some videos are no longer publicly available as the channel owner has deleted them.
The script will also generate the depth and segmentation maps of the images to get the scene context label.
After downloading data, we use WHAM to extract pseudo-labels of the pedestrian movements. Our custom code for running WHAM is at third_party/WHAM and we recommend installing WHAM in a separate python environment following their official instructions. Next, run the follows under the WHAM environment:
bash run_wham.sh $DATA_ROOT
After data preprocessing, the $DATA_ROOT directory should contain the following folders image video depth semantic wham as well as the label files train.pkl val.pkl.