BSUV-Net-2.0

Requirements

1. Python 3.6.9
2. PyTorch 1.3
3. OpenCV 4.0.1
4. tensorboardX 2.2
5. matplotlib

Dataset

Steps for preparing CDNet2014

1. Download the dataset from changedetection.net and unzip the contents in ./dataset/currentFr

2. Download pre-computed empty and recent background frames with FPM images from Google Drive and place the contents in ./dataset

3. In the end, ./dataset folder should have the following subfolders: currentFr, currentFrFpm, emptyBg, emptyBgFpm, recentBg, recentBgFpm.

Cross-validation

1. Run python train.py --set_number <k> for <k> = 1, 2, 3 and 4 to compute the results for each fold. This code will save the results to log.csv.

2. Follow the steps in notebooks/crossvalidation.ipynb to analyze cross-validation results.

Visualization of Spatio-Temporal Data Augmentations

Follow the steps in notebooks/visualization.ipynb to visualize spatio-temporal data augmentations.

Inference on unseen video

Use this repo for inference: https://github.com/ozantezcan/BSUV-Net-inference

Training and Cross-Validation with other datasets.

Change ./configs/data_config.py and ./configs/full_cv_config.py for training BSUV-Net 2.0 with different datasets.