Multi-Object Tracking



Recently, long term object tracking took the attention of researchers. Object tracking in aerial datasets (taken by drones) are also an important area where the data is typically acquired by the onboard cameras. In such datasets, it is important to track objects of interest for long durations visually. An important and distinct problem, (when compared to the ground taken images/videos), is the tracking of the tiny objects accurately over long durations. This challenge in SkyDATA aims to help researchers dealing with multiple object tracking tasks. The goal in multiple object tracking (MOT) is correlating the objects from one video frame to another correctly (i.e., the aim is solving the data association problem). In SkyDATA, we particularly focus on tracking small and tiny objects. Our small and tiny object definitions can be found on the dataset page. Therefore, we also call this challenge the TinyMOT challenge. TinyMOT challenge differs from video instance segmentation (or video instance detection) tasks as it mainly focuses on the association problem. For evaluating the performance of the tracking algorithms, we use similar metrics as used in the MOT challenge. In particular, we compute the following metrics:


We require a single JSON file for the submission. The file format for the TinyMOT challenge is given below: