Multi_Resolution_Rescored_ByteTrack

The code in this repository is based on the code from YOLOX, YOLOVand ByteTrack.
We are happy to announce that our paper Multi-resolution Rescored ByteTrack for Video Object Detection on Ultra-low-power Embedded Systems has been accepted at the Embedded Vision Workshop at the CVPR conference.
In this repo the code to reproduce our results with MR2ByteTrack on the YOLOXS network. To do so you will need to:

• Install the dependencies you can either:
  • create a new environment using conda with the provided .yaml file conda -conda env create -f Multiresolution_ByteTrack.yml
  • install the dependencies via pip with pip install -r requirements.txt we recoomend an environment with at least python version 3.8.16 installed later version should work.
• Download the weights of the YOLOXs network
• Download the ILSVRC2015 VID dataset from IMAGENET
  • unzip the dataset and keep the structure of the folder unchanged
• Modify the yolo_base_multisize.py file in the Experiments folder so that it cointains the directories for the ILSVRC2015 VID dataset, specifically:
  • change self.val_dat_dir with the path to the Data VID folder contained in the ILSVRC2015 VID dataset folder
  • change self.val_ann_dir with the path to the Annotation VID folder contained in the ILSVRC2015 VID dataset folder
    To reproduce the experiment with MR2-ByteTrack and the YOLOXS network run the command:

   python eval_multires.py -f=./BYTE_RESCORE/Experiments/BYTE_YOLOXS_multisize.py -c=path/to/the/yoloxs/weights

  the -f option indicates the experiment that you are running and -c is used for loading pretrained weights from checkpoint. For an explanation of all the parameters see the file yolo_base_multisize.py.