How to interpret predicted KITTI label output

traveller59 / second.pytorch

SECOND for KITTI/NuScenes object detection

MIT License

1.71k stars 721 forks source link

How to interpret predicted KITTI label output #160

Closed Hunkzer closed 5 years ago

Hunkzer commented 5 years ago

Hello!

I implemented the PointPillars solution and since it depends on the SECOND code base, I'm asking for some advice here:

I compared the labels from the evaluation against ground truth labels from the official KITTI data set:

Screenshot from 2019-04-16 10-07-40

Why are there so much more cars in the predicted file? Is my implementation working as intended? I appreciate your help :)

~Hunkzer

traveller59 commented 5 years ago

these cars are low-score cars and they won't harm performance. If your detector have too much false negatives, you can decrease the score threshold to decrease false negatives (increase false positives) and increase KITTI AP. In addition, I have trained a xyres_16 model in this project but only get 77.4 val 3D AP (77.9 in their paper).

Hunkzer commented 5 years ago

Thank you for your very fast response!