reprodcing issue

[jtArthur9]

Hi!
Congratulations to your great work!
I've been trying to reproduce your work, but I'm running into some issues. Specifically, when I use KITTI 05, 06, 07, and 09 sequences as the training set, the average precision of loop cllosure detection (using Protocol 1) on KITTI 00 sequence and KITTI-360 09 sequence is basically consistent with the paper (KITTI 00: 0.97, KITTI-360 09: 0.83), but the average precision obtained on KITTI 08 sequence and KITTI-360 02 sequence are poor, about 30 percentage points lower than in the paper. I'm confused about this, can you provide some insight into what might be causing this issue? Thank you so much. The GPUs I used for training were 2 NVIDIA GTX2080Ti, batch_size was 2, epoch was 150 and point_cloud_augmentation was enabled.

Ubuntu18.04, cuda: 10.2, pcdet: 0.6.0, open3d: 0.14.1, spconv: 1.2.1, pytorch: 1.9.0，faiss：1.7.4