Custom data and architecture about ab3dmot

[xinshuoweng]

1. Yes. It does not matter what method you are using for detection. As long as you get the detection file, you can use our tracking method to do the association.
2. I am not sure what you mean here. If the detection part is running at 30 FPS and the tracking part is 214.9 FPS, then the overall detection and tracking system should be able to achieve around 30 FPS.
3. Yes. Our tracking method is class-agnostic. As long as you get the detection for other classes, the tracking will work.

[abhigoku10]

@xinshuoweng thanks for the response , i have elborated the first point

1. i have model which is trained on BEV data which will give bounding box output and class confidence score , can we use this for tracking using ur source code because the input for ur source code has 2D , 3D orientation values so how to obtain those values . any thoughts on this

[xinshuoweng]

If you do BEV bounding box prediction, I guess you already have the 3D orientation values (orientation around the yaw axis). I guess what you do not have is the 2D bounding box value, 2D camera orientation, also height in 3D size and z in the 3D center. You can simply fill a dummy value for all of these values you are missing, e.g., 1 for all of them. Then I guess you should be fine. :)

[abhigoku10]

@xinshuoweng thanks for the input now its working completely