search

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).
Apache License 2.0
765 stars 96 forks source link

How to choose the score threshold of the seed prediction for fsd++? #163

Closed eagle-chase closed 6 months ago

eagle-chase commented 9 months ago

I have trained a fsdv2 model to generate seed predictions for fsd++, but I don't know how to choose the threshold of detection score. I just chose a score with better visualization results. Is it better to increase the threshold to reduce false positives, or to lower the threshold to increase the recall rate？Have you ever do some experiments？

eagle-chase commented 9 months ago

And I found that sometimes detecting the direction of the target can be reversed (error=pi) on my own dataset. Is there a targeted solution to this problem , or can it only be solved by expanding the data scale.

Abyssaledge commented 9 months ago

It is hard to say what threshold is better, while I prefer to slightly a higher threshold to avoid too many false positives. Is the heading flip frequent? I guess occasionally flip can be fixed by tracking.