The AP on KITTI val set

[yjy4231]

May I ask if the result of your KITTI val set is the result of the last epoch selected?
I found that your results on the KITTI val set are much lower than those on the KITTI test set, and when I reproduced them, I found that the highest KITTI AP | 3D from 52 epoch can reach 22.50/17.61/16.94, and from then on, the AP gradually decreases to
18.22/14.17/13.28.

2. If the result of your KITTI test set is also selecting the last epoch and submitting it to the official benchmark?

[cskkxjk]

1.We evaluate the last 10 epoch ckpts on the KITTI val and report the highest one.
It is possible the early epoch model can achieve better performance for not overfitting since KITTI is not a very large dataset.
Also the performance gap between epochs (epoch 16-20) become small on Waymo (a larger dataset) based on our evaluation.
2. We choose 3 ckpts for (random from 51-53, 54-56, 57-60) epochs and submit to the KITTI website and report the best.