Why is the KLT VPI limited to 128 bounding boxes?

Robotics & Edge Computing Jetson & Embedded Systems Jetson AGX Xavier
1

Hey!

I’m using the KLT tracker in a different way. Since the KLT tracker can only track bounding boxes that max size 64x64 (for some reason…), I subdivide my image using a max box size and stride to split into multiple smaller bounding boxes. In my use case, I can have 128+ trackers running. The problem is that the KLT tracker is limited to 128 boxes for CUDA/CPU and 64 for PVA. Is there a way I can remove the restriction on the CUDA/CPU side and even the PVA? I want to be able to try to track at least 500 small objects over the image. Can somebody at Nvidia compile a patched version with the restriction removed for me? I find the GPU/CPU usage pretty slow for using the tracker so it would be a nice to have.

3

Hi,

In our document, we only find a limitation for PVA.
https://docs.nvidia.com/vpi/algo_klt_tracker.html#autotoc_md46

Have you tried to enlarge the capacity of VPIArrayData to see if it works?

VPIArrayData data_bboxes;
...
data_bboxes.capacity    = 128;
...

Thanks.