3D objects tracking

[goktugyildirim]

How can I evaluate my 3d tracker? providing 3d euclidean distance between objects and hypothesis is enough to calculating mota and motp? is there any example to do this?

[cheind]

Hey,

short disclaimer: I am not aware of the state-of-the-art in 3d tracking nor common benchmarks/metrics.

Most of the current metrics computed in mot-metrics rely on solving some kind of assignment problem that is commonly formulated as finding the minimum cost assignment between predictions (M) and ground truth objects (N), subject to some conditions. The costs are given by an MxN real matrix whose entry ij represents the cost of assigning prediction i to ground truth j in the given frame. These cost of entry ij is commonly computed by intersection over union ratio of prediction and ground truth bboxes. So, one generalization to 3D extends the intersection over union metric to 3D, which leads to comparing the intersection over union volume of two bounding boxes in 3d. The remaining code could stay the same.

[goktugyildirim]

If I directly use the distance between the center of the predicted 3D box and the ground truth 3D box as a cost instead of the intersection over union ratio, can I evaluate the 3D tracker performance correctly? Also, this may be a good example https://github.com/AlienCat-K/3D-IoU-Python/blob/master/3D-IoU-Python.py