Questions about the effect of point cloud registration

[GravityTnT]

Excellent work！I achieved satisfactory registration results using the RAP project in some angularly-defined map scenes. However, I failed to obtain excellent registration results on some datasets with complex indoor scenes. Could it be that my configurations are incorrect?
The dataset I used is 3DMatch/kitchen.

Some of the parameters I set are as follows：

[GravityTnT]

Follow up. I used 4 point clouds from the 3DSparseMatch/kitchen dataset with a similar scenario and achieved good registration results. So I'm curious as to why the above problem occurred.

[YuePanEdward]

Thanks for your interest in our project. You may need to turn on the option: Apply frame transform (for 3DMatch-like data with Z-axis pointing forward). Besides, I will soon update a model that would have better performance with a more flexible rotation setup.

[GravityTnT]

Thanks for your interest in our project. You may need to turn on the option: Apply frame transform (for 3DMatch-like data with Z-axis pointing forward). Besides, I will soon update a model that would have better performance with a more flexible rotation setup.

Thank you for your prompt reply. I will continue to make attempts on the 3Dmatch dataset. Could you please tell me if this update will also include the training code? I'm really looking forward to your next steps.

[YuePanEdward]

Yes, I will also update the training code and some example training data soon.