Track time to loong

[Lorenzuou]

I have a tracker + yolo_nas setup, at first the predict function from yolo_nas super_gradients was to slow, so i changed it and now it runs at 30ms

but the tracker is running at 100ms-300ms, all deepSort, strongSort and botSort are doing the same time in average. This speed makes my stream application to slow

how can i speed up the tracking? (yes, i am using at gpu enviroment)

[mikel-brostrom]

You could use bytetrack or ocsort instead, which are ReID model free tracking methods

[Lorenzuou]

Yes, but, is the latency I mentioned for tracking methods with models that slow? I am using a 3060 to run it with 640x640 images

[mikel-brostrom]

yolo_nas is SLOW. Their inference methods have large overheads. For some reason they don't intend to fix this.

[Lorenzuou]

I change it to get the raw output from usual inference and it got better, funny thing is they put the latency ate their page at 7 ms

[mikel-brostrom]

I thinks this is only for their quantized onnx models. Their .pt inference is poorly optimized