Iterates through the detections per frame, and assigns detections from the next few frames to tracklets based on distance and other losses using the Hungarian algorithm.
Vaguely based on SORT, except with more engineering (so not as simple), and a few global operations (like smoothing; so not online). So, I guess it's just RT. It's still quite fast. Mostly just manual tweakable parameters. No learned weights. Simply uses the detections as-is and hopes.
This tracker was initially designed for a different project wherein the detections were high quality and there were very few self-occlusions. It was adapted for this challenge over the course of one day (including all hyper-parameter tuning). Let's see how this goes.
The biggest issue to note is the lack of appearance features. This means that it can't tell whether a box belongs to the same person if they haven't been detected for more than a few frames, or when it starts tracking someone else. I think that a re-identification method could be used to join the tracks produced by this algorithm.
Note: The reported "Runtime" is the total time taken to evaluate on both the training and the test set (I'm not sure if it's asking for just test set time, so I put the more conservative value).
3.7GHz, 1 Core
June 12, 2019 (1 month ago)
June 12, 2019 at 12:55:51 CET
Project page / code:
|33.9||79.9||3.5||7.1 %||38.8 %||15,477||350,754||4,020||4,748||3.7GHz, 1 Core||Public|
|IDF1||ID Precision||ID Recall|