search

haoyu94 / RoITr

Rotation-Invariant Transformer for Point Cloud Matching
MIT License
111 stars 13 forks source link

Questions about performance with other partially overlapping data #15

Open CorgiBoyG opened 3 months ago

CorgiBoyG commented 3 months ago

Thanks for making such a great work. Recently, I tried to train RegFormer on ModelNet, ModelLoNet, Kitti, but the results are not so good. For example, I trained on ModelNet with no modifications to the network, optimizer, and other parameters, and immediately after training for 150 epochs, I evaluated it, and the final rre calculated for the validation set was 42. Could you please tell me what modifications I need to make, e.g., parameters, etc.? Can you please provide some suggestions to help me, it is very important for me, thanks! Thanks again for the superior work you have done. Sincerely looking forward to your reply!

k794102577 commented 3 months ago

I fixed the kitti problem, but not the modelnet problem

k794102577 commented 3 months ago

For modelnet, did you test it on a clean, complete point cloud, and how did it work

CorgiBoyG commented 3 months ago

I fixed the kitti problem, but not the modelnet problem

How did you resolve this and can you offer some advice?

CorgiBoyG commented 3 months ago

For modelnet, did you test it on a clean, complete point cloud, and how did it work

I have only tried modelnet under partial overlap, following the rpmnet protocol for partial overlap data generation. Found it didn't work well, rre was only 42.

k794102577 commented 2 months ago

My suggestion is to adapt the entire framework to geotransformer parameters and then load kitti data using predator preprocessed code,it is not difficult

k794102577 commented 2 months ago

For the modelnet dataset, I can also trainning smoothly

CorgiBoyG commented 2 months ago

For the modelnet dataset, I can also trainning smoothly

Okay, could you tell me how you adjusted it, please. I want to use roitr as a comparison algorithm, not an improved optimization based on Roitr. If it's convenient for you, could you guide me in detail what I should change. I tried to change the parameters of roitr yesterday against GeoTransformer, but the training remains the same!

CorgiBoyG commented 2 months ago

For the modelnet dataset, I can also trainning smoothly

Excuse me. Can you help me please, I'm only using it for comparison algorithm experiments, not plagiarizing. This is very important to me, thanks. Looking forward to your reply.

k794102577 commented 2 months ago

First of all, we also used the preprocessing code of predator for the modelnet data set, which was verified and found that the framework could be trained smoothly under crop, jitter and clean conditions, and the test results were found to be normal. After the training, I will upload the screenshot

CorgiBoyG commented 1 month ago

First of all, we also used the preprocessing code of predator for the modelnet data set, which was verified and found that the framework could be trained smoothly under crop, jitter and clean conditions, and the test results were found to be normal. After the training, I will upload the screenshot

May I ask how your subsequent experiments have worked out and if you are getting good performance?

CorgiBoyG commented 2 weeks ago

My suggestion is to adapt the entire framework to geotransformer parameters and then load kitti data using predator preprocessed code,it is not difficult

If it is convenient for you, is it possible to provide the modified code? I promise to use it only for experimental comparison testing.

k794102577 commented 2 weeks ago

You can comment out the data enhancement code without affecting the final comparison test of the experiment