Skip to content

Latest commit

 

History

History
67 lines (59 loc) · 2.81 KB

README.md

File metadata and controls

67 lines (59 loc) · 2.81 KB

Pre-processing Stage

To properly perform the pre-processing stage, please change your directory to:

cd path-to-directory/Interpretable_VAD/

1. Object Detection

Our object detector outputs are provided here. Set up the bounding boxes by placing the corresponding files in the following folders:

  • All files for Ped2 should be placed in: ./data/ped2
  • All files for Avenue should be placed in: ./data/avenue
  • All files for ShanghaiTech should be placed in: ./data/shanghaitech

This section describes how to prepare the object detector to extract bounding boxes:

Please install the Detectron2 library by executing the following commands:

python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

Then download the ResNet50-FPN weights by executing:

wget https://dl.fbaipublicfiles.com/detectron2/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl -P pre_processing/checkpoints/

Run the following command to detect all the foreground objects.

python pre_processing/bboxes.py [--dataset_name] [Optional: --train]

E.g., In order to extract all train objects from Ped2:

python pre_processing/bboxes.py --dataset_name=ped2 --train

This will save the results to ./data/ped2/ped2_bboxes_train.npy, where each item contains all the bounding boxes in a single video frame.

In order to extract all test objects from Ped2:

python pre_processing/bboxes.py --dataset_name=ped2

This will save the results to ./data/ped2/ped2_bboxes_test.npy, where each item contains all the bounding boxes in a single video frame.

2. Optical Flow

We extract optical flows in videos using use FlowNet2.0. In order to install FlowNet2 please execute the following commands:

cd pre_processing
bash install_flownet2.sh
cd ..

First, download the pre-trained FlowNet2 weights (i.e., FlowNet2_checkpoint.pth.tar) from here and place it in Interpretable_VAD/pre_processing/checkpoints/. Now, Run the following command to estimate all the optical flows:

python pre_processing/flows.py [--dataset_name] [Optional: --train]

E.g., In order to extract flows from Ped2 train frames:

python pre_processing/flows.py --dataset_name=ped2 --train

This will save the results to ./data/ped2/training/flows/, where each item contains all the bounding boxes in a single video frame.

In order to extract flows from Ped2 test frames:

python pre_processing/flows.py --dataset_name=ped2

This will save the results to ./data/ped2/testing/flows/, where each item contains all the bounding boxes in a single video frame.