search

qinzheng93 / GeoTransformer

[CVPR2022] Geometric Transformer for Fast and Robust Point Cloud Registration
MIT License
636 stars 67 forks source link

The training model is not effective #48

Open awdx551 opened 1 year ago

awdx551 commented 1 year ago

Thank you for your excellent work. When I use the LGR training model without changing the code, the result on 3DMatch (about 91.0%) was still quite different from the model you provided (92.5%)

qinzheng93 commented 1 year ago

This could be caused by two reasons. First, different random numbers during training can affect the convergence, especially with a batch size of 1.. Second, there is a gap between correspondence and registration. As we only supervise the correspondence, we expect to achieve better registration results with better correspondences, but it is not always the case. You can try the model with best RR on the validation set. And are the IR and the FMR consistent with the results of the pretrained model?

awdx551 commented 1 year ago

Thank you for your reply! I did not change the random numbers for training, and I did use the model with the best RR on the validation set. As you said, IR and FMR have the same or even better effect than the pretrained model. In theory, RR should not be different from the pretrained, which is what I don't understand.

qinzheng93 commented 1 year ago

It is somehow weird. There could still be something in the model that is not deterministic during training. I will have a look at it in the future.