track_video

Each detection instance should contain (x1, y1, x2, y2, score [, cls]). While the class prediction is not required, it is still useful to save for custom filtering. Otherwise, each track additionally keeps track of the moving centers, sizes

Option Semantics

Before: necessary to know whether tracking is enabled

Tracking	OFF	ON
--det-all	dets as tracks	track all
person only	dets as tracks	track persons

After: explicit detections and person tracks

Tracking	OFF	ON
--det-all	all dets, no tracks	obj dets, person tracks
person only	person dets, no tracks	no dets, person tracks

Implementations

DeepSort

• Detection: (x, y, w, h, score, features)
• Track: (id, mean, cov, features,...)

Traction

• Detection: (C, x1, y1, x2, y2, p, f, di)
• Track: (id, C, cx, cy, w, h, history)

p: detection probability or score
f: object features
di: detection index w.r.t. detection frame
history: last few detections

Datasets

• Wildtrack