search

IRMVLab / RegFormer

[ICCV2023]RegFormer: An Efficient Projection-Aware Transformer Network for Large-Scale Point Cloud Registration
108 stars 5 forks source link

Question about the code used for Nuscenes training having a negative loss when trained on modelnet40 #14

Open CorgiBoyG opened 4 months ago

CorgiBoyG commented 4 months ago

Thanks for doing such a great job and for providing the network code for Nuscenes training. I noticed that the projection parameters in the code for loading the dataset for Nuscenes are self-generated. I then referenced this for the modelnet40 dataset, where the overlap handling is referenced to the RPMNet protocol. Regarding the projection parameters, I kept them the same. The result is that after epoch about 2~4, the loss is negative. What should I do and could you provide some good suggestions to help me please. Looking forward to your reply and thanks again for doing such a good job.

yuxuanfanOrion commented 4 months ago

Hi there, Can you reproduce the results in the article?

liujiuming123 commented 3 months ago

Hi, I don't know whether our outdoor registration model can adapt well to the object-level or indoor scenes, because the projection parameters are set according to the physical LiDAR's resolution and range. Like for KITTI and Nuscenes, that have totally different vertical coverage and beam numbers (KITTI 64 beam NuScenes 32 beam). So for these mechanical scanning LiDAR, its projection is highly corelated to the sensor. But for ModelNet40, maybe there lack such physical constraints.