I tried Optical Flow SDK V2.0. The tracking result is good in non-occlusion scene.
I changed the option : Detection skip intervalle and I don’t understand why if I set this value more than 3, the tracking failed. I mean, if I use this command :
NvOFTSample.exe -i TownCentre.mp4 -e yolov3.trt -sI 3
The tracker do great job. But if I put the parameter to 4 or 10, in the video result, the objects tracked changed its ID and sometimes the bounding box disappeared and come back again (detection frame).
I want to use Optical Flow SDK to track object and avoid doing detection each frame. But I want to skip more than 3 frames to deal more devices (I know that the accurancy could decrease by increasing skip frame detection). But I expected that Optical flow can tracker without loss the object in more than 3 frames without detection.
Suppose that if I skip 3 frames for detection, If the detection failed to detect an object, I expected that tracker continue to track the object to compensate a failure of a detection.
Is it because the sample was coded to erase objects which has not been detected more than 3 frames ?
I read the documentation and it talk about skipping frame (0 <= K <= 4) to let GPU to do intensive task in other channel, don’t know if the sample is an implementation of the sample in documentation.