Otteri/point-pillars

Confidences below ScoreThreshold

Opened this issue · 0 comments

It appears that some detection confidences can be really small, as an example: 1.66316e-07

These detections seem good though and there is no reason to filter these out. The close to zero confidence detections do not appear when threshold is high, but are more frequent with lower threshold value --> threshold value has effect to these. Seems like that these are proper detections with confidence above the treshold, but the confidence value is just incorrect.

The low value is already visible in postprocess.cu. One may check host_filtered_scores[keep_inds[i]] variable values for example. This value is calculated on GPU and then copied to host. So, it seems that error happens on GPU.

My guess is that the issue comes from bad type conversion and this bad type conversion exists, because pytorch model uses higher floating point precision than the the final TRT model.

There is a chance that this issue disappears when libraries are updated.