Moving-Camera Background Subtraction Network for Autonomous Driving

MBSNet

This project aims at detecting salient background which in return provides regions of moving objects referred to as the foreground. This is an alternative approach for object detection.

Software Installment

The code is implemented on Python 3.7.

• PyTorch 1.6.0
• pydensecrf
• opencv 4.4.0.40
• tmdq 4.48.2

Training

1. Download datasets and add local dir for training and validation images and labels.
2. Modify parser.add_argument in train.py according to your requirements. Input arguments include:

• --image_dir
• --label_dir
• --batch_size
• --backbone
• --start_epoch
• --Deconvolution

Descriptions about above arguments are available on train.py.

3. Train MBSNet

Inference

Inference datasets are also available. Please assign a GPU for inference if available. speed_analysis.py is written for testing trained MBSNet inferencing speed.

1. Add local testing images and labels dir.
2. Modify parser.add_argument in test.py according to your requirements. Besides those input arguments listed above train.py, test.py also includes:

• --type
• --class_
• --epoch
• --use_crf
• --crf_num

3. Trained MBSNet under Xception39 basemodel is available OneDrive.

Disclaimer

This project is an ongoing project. Future work will be released. Feel free to open an issue if you get stuck anywhere.