In Conjunction with the Conference on Computer Vision and Pattern Recognition, CVPR 2020. The workshop will happen live via zoom and also be broadcasted here.

For this 5th edition of our Benchmarking Multi-Target Tracking (MOTChallenge) Workshop, we want to push the limits of tracking by focusing on tracking at pixel accurate level. To achieve this we introduce the MOTS20 (Multi-Object Tracking and Segmentation 2020) benchmark. For this we have densely annotated 4 training sequences and 4 challenging test sequences from the challenging MOT17 dataset with pixel accurate segmentation masks. The data features crowded pedestrian scenes with many difficult cases of occlusion.

In order to enourage research towards end-to-end systems which solve the whole MOTS task together, for this challenge the partipants are not required to use "public detections", but can rather use any detector. We also include a challenge track where we provide state-of-the-art segmentation predictions and participants can focus on the tracking aspect without having to perform detection and segmentation.

We also host another challenge track on the KITTI-MOTS dataset, where methods are required to detect, segment and track both cars and people in an autonomous vehicle setting.

Our workshop has three major goals: (i) Analyzing the performance of state-of-the-art detection, tracking, and segmentation algorithms, and their combinations on crowded pedestrian scenes, (ii) opening new Multi-Object Tracking and Segmentation Challenges on 4 the MOTS20 and KITTI MOTS datasets, and (iii) discussing the limitations of current methods for producing accurate segmentation masks under severe occlusion scenarios.

An important part of the workshop will also be dedicated to a discussion among participants on how to improve Multi-Object Tracking and Segmentation evaluation and ideas on how to expand the current benchmark. These discussions in previous editions of the workshop have helped us tremendously in shaping MOTChallenge and significantly contributed to creating a widely used and perhaps the most popular multi-object tracking benchmark in the community.

свернуть