This repository contains the source code of BSUV-Net algorithm as described in the following paper:

BSUV-Net: A Fully-Convolutional Neural Network forBackground Subtraction of Unseen Videos by M. Ozan Tezcan, Prakash Ishwar and Janusz Konrad.

BSUV-Net is a convolutional neural network which uses the concatenation of several images as input. The descriptions of these images can be found in the paper.

Currently source code contains only the trained models and inference code that can be applied to any video. examples folder includes an example video and its background subtraction result using BSUV-Net with and without FPM.

1. Python3 (testen on v3.6)
2. PyTorch (tested on v1.3 and v1.4 with CUDA 10.1 on Ubuntu)
3. OpenCV for Python (tested on version 4.2.0)
4. YACS (pip install yacs should work)

Trained model which uses the RGB channels of empty background, recent background and current frame (9-channel input) can be downloaded at: https://drive.google.com/file/d/1q6a0RJuD54Gq8txw6TKPRRnC5xnSucLR/view?usp=sharing

Trained model which uses the RGB+FPM channels of empty background, recent background and current frame (12-channel input) can be downloaded at: https://drive.google.com/file/d/1ISzZyLDzuRuMnNmrZ3QVJCVeT_3eltDK/view?usp=sharing

Trained weights for BSUV-Net 2.0 (12-channel input) can be downloaded at: https://drive.google.com/file/d/12y2PMK8Ne7G27CI5Vx6hC2qkoF5cJo5I/view?usp=sharing

Trained weights for Fast BSUV-Net 2.0 (12-channel input) can be downloaded at: https://drive.google.com/file/d/12y944z5yPePy2RFPu8V4VQr9IfBhLjUM/view?usp=sharing

1. If you want to to use a model with FPM channel, download HRNet_v2 model files from http://sceneparsing.csail.mit.edu/model/pytorch/ade20k-hrnetv2-c1/ and place them in utils/segmentation/hrnet_v2 folder.
2. Change configs/infer_config.py based on your application.
3. Run python inference.py <vid_in> <vid_out> where <vid_in> is the path to the input video and <vid_out> is the desired output path.