Test KITTI-trained PGD model on nuScenes

[Yibin122]

score_thr = 0.9:

I checked the class labels of those FP predictions and there are not only pedestrians but also cars.
Then I checked the scores and found all of them are > 1. That's why increasing the score threshold does not suppress the bad results.

Why would the score value be greater than 1 (even after Sigmoid)?

[Tai-Wang]

That's because PGD involves a strong uncertainty score on KITTI to suppress those predictions with inaccurate depth estimations. Please refer to the discussion in issue #1024 for more details.

[Tai-Wang]

Close due to inactivity and duplication.