VAE for Video Instance Segementation Tracking
This source code is a personal customized implementation for wild animal datasets based on the proposed idea in Video Instance Segmentation Tracking with a Modified VAE Architecture
- Install dependencies
pip3 install -r requirements.txt
- Videos captured wild animals of different classes.
- Frame size: (720, 1280). It will be resized to (768, 1280) while dataset modul is defined.
There are 5 modules in this implementation:
- dataset.py : defines datasets for the training and evaluation of Mask-RCNN and VAE, reading from wild animals dataset annotation files.
- model.py: contains model classes.
- utils.py: contains related functions.
- evaluation.py: post-processes model's prediction and saves results in .txt file for MOT, MOTS, and COCO evaluation.
- run.py: configures dataset path and properties, and sets up model's mode (trains or uses the trained parameters to get prediction results).