Tools and example for Challenge 3: Tracking Only (MOT+KITTI) of MOTChallenge 2020 (https://motchallenge.net/workshops/bmtt2020/tracking.html)
In this challenge, you're given strong pre-computed detections with segmentation masks and your task is tracking only, i.e. you are only required to sub-select from the given masks, assign these consistent tracking IDs, and assign a score to each selected mask based on which overlapping masks will be combined (the final result must be non-overlapping). Note that you are not allowed to "fill gaps" by creating your own detections/masks. We created the detections using the model Mask R-CNN X152 of Detectron2 and afterwards running refinement net (https://arxiv.org/abs/1807.09190) to improve the mask quality.
Your tracker has to produce a txt format output. Each line has the following format: det_id track_id mask_merge_confidence
Where det_id is the id of the detection in this sequence (i.e. the line number starting from 0 of the detection file) track_id is an id of a track, you have to create these ids yourself and multiple detections can be mapped to the same track id mask_merge_confidence is a float, for overlapping masks, the mask with the higher value will be on top
For an example to create results using a "dummy tracker" and evaluate them, run ./run_demo.sh
For questions please contact Paul Voigtlaender via [email protected]