• ByteTrack
  • Effective and generic association method, tracking by associating almost every detection box instead of only the high score ones
• BYTE
  • First Association
    • Matcing high score detection bounding boxes based on motion similarity and appearance similarity
    • Next, Kalman filter is applied to predict the position of the tracklets of the next frame
    • This process is similar to traditional trackers
  • Second Association
    • The difference of BYTE is considering the low score detection bounding boxes once more
    • Based on the same motion similarity, we match unmatched tracklets and low score boxes such as the occlusion boxes
• Similarity metric: IoU-based (re-id is not supported)
• It is dependent to performance of detection.
• More details
  • https://blog.naver.com/qbxlvnf11/222784457900
  • ByteTrack: Multi-Object Tracking by Associating Every Detection Box
• Pseudo-code of BYTE

• You only look once (YOLO) is one of the the powerful and real-time 1-stage object detection systems
• Improved features compared to yolov2: FPN,shortcut connection, logistic regression etc.
• More details: YOLOv3: An Incremental Improvement

• Identifying objects detected by yolov3

• Train yolov3 model
• Detect image

• Convert yolov3 Pytorch weigths to TensorRT engine
• Real-time inference with yolov3 TensorRT engine

• byte_tracker_config.ini: byte tracker parameters
• yolov3_config.ini: yolov3 model parameters
• train_config.ini: yolov3 train parameters
• tensorrt_config.ini: yolov3 tensorrt parameters

• https://docs.nvidia.com/deeplearning/tensorrt/container-release-notes/rel_20-12.html#rel_20-12

docker pull qbxlvnf11docker/byte_tracker_yolov3:latest

nvidia-docker run -it --name byte_tracker_yolov3 -v {yolo-v3-tensorrt-repository-path}:/workspace/Byte-Tracker-Yolov3 -w /workspace/Byte-Tracker-Yolov3 qbxlvnf11docker/byte_tracker_yolov3:latest bash

• Params: refer to config files and parse_args()

python main.py --mode byte-multi-objects-tracking-images

./create_model_def.sh {class_num} {cfg_name}

./download_weights.sh

• Params: refer to config files and parse_args()

python train.py --mode yolov3-train

• Params: refer to config files and parse_args()

python yolov3_convert_onnx_tensorrt.py --yolov3_config_file_path ./config/yolov3_config.ini --tensorrt_config_file_path ./config/tensorrt_config.ini

• GOT-10k: http://got-10k.aitestunion.com/
• DanceTrack: https://sites.google.com/view/dancetrackmot

• Download COCO2014 dataset

./get_coco_dataset.sh

• Building data json for optimizing yolov3
• In train process, read builded data json file and get train data
• Params: refer to parse_args()

python yolov3_convert_onnx_tensorrt.py --target coco2014 --data_folder_path ./data/train_data/coco --save_folder_path ./data/data_json/coco

• parsing_data_dic['class_format'] = type of class ('name' or 'id')
• parsing_data_dic['label_scale'] = scale of label ('absolute' or 'relative')
• parsing_data_dic['image_list'] = [{'id'-image id, 'image_file_path'-image file path}, ...]
• parsing_data_dic['object_boxes_list'] = [{'image_id'-image id, 'object_box_num'-number of the object per image, 'object_box_id_list'-[object box id, ...], 'object_name_list'-[object name, ...], 'object_box_list'-[[center x, center y, box_width, box_height], ...], 'object_box_size_list'-[object box size, ...], }, ...]
• parsing_data_dic['image_num'] = number of the image
• parsing_data_dic['object_boxes_num'] = [number of the total objects, ...]

@article{ByteTrack,
  title={ByteTrack: Multi-Object Tracking by Associating Every Detection Box},
  author={Yifu Zhang et al.},
  journal = {arXiv},
  year={2021}
}

@article{yolov3,
  title={YOLOv3: An Incremental Improvement},
  author={Redmon, Joseph and Farhadi, Ali},
  journal = {arXiv},
  year={2018}
}

https://github.com/ifzhang/ByteTrack

https://github.com/qbxlvnf11/yolo-v3-tensorrt