This challenge is about Multi-Object Tracking and Segmentation of PEDESTRIANS on the MOTS20 test set.

Downloading the train and test data, leaderboard and submission site is available on the MOTChallenge Website.

Evaluated on the 4 test sequences of MOTS20.

No detections or segmentations given, methods must perform detection, segmentation and tracking.

The same method with the same parameters must be run on all four sequences. No per-sequence tuning is allowed.

In this challenge, the MOTS task is evaluated on the KITTI-MOTS test set for both CAR and PEDESTRIAN classes.

Training and Test data are available, a live leaderboard and the submission site are live. All can be found here.

All entries on the KITTI-MOTS leaderboard between January 1, 2020, and May 30, 2020 will count as challenge entries.

Winners will be determined by the highest sMOTSA score, averaged over the sMOTSA for CARS and PEDESTRIANS.

No detections or segmentations given, methods must perform detection, segmentation and tracking.

Methods are allowed to use available stereo, LiDAR, and GPS if they wish.

The method and parameters for CAR and PEDESTRIAN class can be different, but within the same class the method must be the same with the same parameters for each video.

Evaluated on both the MOTS20 test set and the KITTI-MOTS test set for both CAR and PEDESTRIAN classes.

This is a tracking-only challenge. Competitors must use the given segmentation detections, and are only required to sub-select from the given masks, assign these consistent tracking IDs, and rank them based on which should be on top when masks intersect.

The tracking method must be EXACTLY the same for MOTSChallenge and KITTI-MOTS. E.g. there may be no dataset specific or video specific parameters.

Methods are NOT allowed to use stereo, LiDAR, or GPS on KITTI. Monocular image input only.

Here, you can find the detections and tools to train and produce the desired output.

Note that we pre-computed detections for all sequences of MOTS-17, but for the challenge you only have to provide results of the MOTS20 test set (i.e. sequences 1, 6, 7 and 12), and of the whole KITTI MOTS test set.

Further you are not allowed to run additional detectors.

Challengers are only allowed to submit ONCE, by emailing their submission results, in the .txt format required (see tools) to voigtlaender@vision.rwth-aachen.de.

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 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.