Köferl F, Link J, Eskofier B (2020)
Publication Language: English
Publication Type: Conference contribution, other
Publication year: 2020
Event location: Seattle, WA, USA (Virtual)
URI: https://motchallenge.net/workshops/bmtt2020/papers/Application_and_Adaptations_of_SORT_on_MOTS20.pdf
Multiple object tracking and segmentation (MOTS) on
monocular images using object detectors without any end-to-end learning of the tracking step has been a common
problem historically. Including the posterior of the object detector into the tracking step proves to be difficult, because
features from the detection step are reduced to only segmentation mask, object probability, and class information.
Based on this, solving tasks like combining of segmentation
masks and the actual tracking step is still the main challenge. We adapt an existing simple online tracking method
(SORT) based on bounding boxes. The tracking process
predicts trajectory using a Kalman filter and matches tracks
to detections using a simple IOU metric. The sMOTSA score
on the test set of KITTI-MOTS are 64.1 (cars), 54.5 (pedestrian) and on the test set of MOTS20 is 56.8 (pedestrian).
APA:
Köferl, F., Link, J., & Eskofier, B. (2020). Application of SORT on Multi-Object Tracking and Segmentation. In Proceedings of the Conference on Computer Vision and Pattern Recognition; 5th BMTT MOTChallenge Workshop: Multi-Object Tracking and Segmentation. Seattle, WA, USA (Virtual).
MLA:
Köferl, Franz, Johannes Link, and Björn Eskofier. "Application of SORT on Multi-Object Tracking and Segmentation." Proceedings of the Conference on Computer Vision and Pattern Recognition; 5th BMTT MOTChallenge Workshop: Multi-Object Tracking and Segmentation, Seattle, WA, USA (Virtual) 2020.
BibTeX: Download