Overview

This repo contains the accompanying code for our paper Motion Extrapolation for Video Object Detection Best Paper Runner Up at the Robust Video Scene Understanding: Tracking and Video Segmentation CVPR 2021 workshop.

WARNING: There is active work being done on this repository. We make no promises about its stability or backwards compatibility for the time being.

Some Differences From the Paper

Since the time of submitting the referenced paper, additional work has been done.

Changes made are:

• MOT16 as an additional evaluation dataset
• global frame compensation implemented for video with a lot of camera movement (like MOT16 sequences 05 and 13)
• ablation options added for removing global frame compensation, aggregation function, multi-processing capabilities, and propagation at large.
• additional optical flow computation types: Robust Local Optical Flow and Dense Pyramid Lucas-Kanade (unstable)

Citation

If you find this work useful, consider citing us:

@misc{true2021motion,
      title={Motion Vector Extrapolation for Video Object Detection}, 
      author={Julian True and Naimul Khan},
      year={2021},
      eprint={2104.08918},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}