Insta-Track: Instance Segmentation Assisted Tracking by Detection

In computer vision, multi-object tracking (MOT) is essential for applications ranging from surveillance to autonomous vehicles. Traditional MOT approaches utilize motion and appearance-based tracking techniques, often employing amodal bounding boxes, which are common in benchmark datasets. Amodal bounding boxes, while stabilizing box size and location for motion-based tracking, can lead to inaccuracies in appearance-based tracking by including non-visible object parts. This study presents InstaTrack, an innovative approach that improves appearance-based tracking through instance segmentation. Trained with contrastive loss and non-negativity enforcements, InstaTrack effectively produces robust appearance features. It calculates coverage values to assess object visibility, thus optimizing the use of appearance features. InstaTrack integrates seamlessly with existing systems, requires no extra datasets, and enhances scene analysis with detailed instance segmentation. Furthermore, rigorous evaluations on datasets with amodal bounding boxes like MOT17, MOT20, and DanceTrack confirm InstaTrack's accuracy and efficiency, showcasing its potential to advance tracking techniques.