Question about the GPU utilization when global tracking

[zhoutianyang2002]

Dear author,

When I use VHAP to track the nersemblev2 dataset, I found that the GPU utilization is 0% during the global optimization, and the CPU utilization is very high(almost occupied all CPU cores of the server). However, during the sequential_tracking stage, the GPU utilization is normal(50%~75% in RTX3090). So it is very strange.

Is this phenomenon normal? I do worry about that there is some potential bugs in the code, but I couldn't find it now.

Thank you very much! Best wishes!

[ShenhanQian]

Hi Tianyang,

you could check if the bottleneck is at data loading. Basically, sequential tracking loads a frame and optimize for fifty iterations, while global tracking loads a new frame for every iteration. If your IO speed is low, GPU utilization will appear very low. To fix this, you could try higher number of workers for the dataloader, or try moving your data to faster storage.

Best,
Shenhan