This repository demonstrates construction of a Single Object Tracker (SOT) using the detectron2 framework. We mainly reuse the Anchor generation part and rewrite the RPN network to suit Object tracking.
The code is based on the SiamFC, SiamRPN, SiamRPN++ papers. Credit also goes to the pysot repository which is an excellent resource for single object tracking in general (doesn't use the detectron2 framework)
We use the COCO2017 dataset to train. We prepare the dataset in the same way as pysot. Refer tp their data preparation here
To train, adjust the COCODataset init parameters and simply run using
python dsot/train.py
Training for 20 epochs already results ina good model (refer video), training futher should result in better results. Additionally data from other datasets like Youtube Video dataste could be added to futher improve accuracy
- Move config to yaml
- Add inferencer from private repo
- Add improvements from SiamRPN++ to improve accuracy