Bird tracking using ByteTrack

This repo contains tools and experiment files used for tracking birds in a cage using ByteTrack.

ByteTrack's is codebase is based on YOLOX object detector, and for this reason, YOLOX pretrained weights can be used for tracking with ByteTrack.

Dataset converter

Under tools there's convert_dataset.py that can be used to convert dataset from COCO format to ByteTrack trainable format.

Install requirements:

python3 -m pip install pycocotools

Example:

python3 convert_dataset.py \
         --input birds/annotations/annotations_val.json \
         --output birds/annotations/annotations_valo_preprocessed.json \
         --data-dir data \
         --ground-truth-output ByteTrack/datasets/mot/train/data/gt/gt.txt \
         --category-as-track_id \
         --coco-categories \
         --target-category 16

Parameters:

• --input: the input COCO file
• --output: the output COCO file
• --data-dir: fix the image file paths in the JSON using the provided directory
• --ground-truth-output: the path of ByteTrack compatible ground truth file that can be used for validation
• --category-as-track_id: if set, COCO category_id will be used as track_id. It can be used when the different moving objects are categorized separately.
• --coco-categories: if set, include COCO categories in the output file
• --target-category: set the annotations' category to this category (16 is the bird)

Browsing the dataset

You can explore the dataset using tools/browse_dataset.py which uses FiftyOne.

Install requirements:

python3 -m pip install pandas fiftyone

Example:

python3 browse_dataset.py -i birds/annotations/annotations_valo_preprocessed.json

Experiment files

The exps directory contains the experiment files which can be used for ByteTrack (and YOLOX) training, evaluation and tracking.

They contain model parameters and dataloading, training and evaluation configs.

• yolox_s_bird.py: for training YOLOX with small network
• yolox_x_bird.py: for training YOLOX with larger network
• bytetrack_x_bird.py: for training ByteTrack with larger network

NOTE: if you use ByteTrack with pretrained YOLOX weight, use the weights from YOLOX 0.1.0 release (git tag), because ByteTrack is currently based on that revision.