Questions about the keypoint numbers and loss functions on KITTI dataset

[SUZhaoyu]

Hi, thanks for the great work. I came across a problem when trying to reproduce the performance on KITTI dataset. I tried different combinations of keypoint numbers (from 64 to 1024) and loss functions (both circle_loss and desc_loss). However, none of my test was able to converge given more than 256 keypoints. Can I have any suggestions on this? Cuz I notice in the paper, the model is able to train even with 5000 keypoints.

[SUZhaoyu]

UPDATE: with all the setup mentioned above, the best success rate I am able to achieve is 95.13%, with RTE=0.25 and RRE=0.96, which leaves a large margin from the paper. I checkout from the previous commit and trained with the default parameters (1024 keypoints with desc_loss), but the model still refuse to converge.