the-learning-and-vision-atelier-lava/buffer

[CVPR 2023] BUFFER: Balancing Accuracy, Efficiency, and Generalizability in Point Cloud Registration

buffer's Issues

How to generate an overlap of a self-made training set

Hi, aosheng
I am a novice, I want to make my own data set, I found that overlap is used in 3dmatch, how should I generate the overlap of the self-made training set.
thanks!

test on my own data

Hello, thank you for your fantastic job!
My question is, whether there is any API to test on my own data? I mean that the input of the API is source point cloud and target point cloud.
Appreciated!

steps to register 3rd-party rigid point cloud parts?

Hi authors,

great work on the point cloud registration! I am very new to this work, but I have a few point clouds in .xyz format with relative start position x/y of each other. I would like to try this package and evaluate how well in registering those parts.

• How to input some 3rd party .xyz file and use existing trained model (e.g. ThreeDMatch's best.pth) for a registration?
  I see there is config.py file where I can set the .dataset and test but not sure exactly how to do it, as my files are not the same format as the existing datasets.

• is there any issues I need to process before processing?
  for example, my dataset would be shape of about 5mx5m real world size with a few million points each part. how do I setup the voxel_size_0 and voxel_size_1 in the config.py file?
  I see existing datasets may have attributes like relt_pose, which does not have in my dataset, will that be an issue?
  any other issues I may face to not give me an appropriate result?

• narrow down with start position x/y. our data is measured with roughly accurate start position x/y to align the parts, but no relt_pose exact file as in the .ply files in the database . Such basic align will narrow down the further registration. How can I input such start position x/y and direction for the processing?

• noise removal. My data is quite noisy, and I see there is some post processing in this package. I am wondering is there some step I should modify to better remove the noise?

• does this software package treat rigid and non-rigid differently? do I need any modification in the code?

thank you and really excited to see this great work!

CZ

Negative numbers appeared when I was training with 3dmatch dataset, is this normal?

details as following:：
Epoch: 3 [10400/33532] data_time: 0.03s model_time: 0.10s
ref_loss: 0.000000
ref_error: 0.000000
desc_loss: 0.000000
desc_acc: 52.416992
eqv_loss: 0.000000
eqv_acc: 0.000000
det_loss: -94.586985
match_loss: 0.000000

Simple inference demo?

Hi again,

I'm wondering if you could provide a simple inference demo on a single point cloud pair.
Right now it seems like I have to use the functions defined in the dataset dataloaders to convert the pointclouds to the input format of the network. Could this perhaps be included as a method in the network itself?
I think this would greatly simplify using BUFFER for downstream tasks.

Can the BUFFER model be tested on the ModelNet40 dataset?

Thanks for your sharing! I tried to apply ModelNet40 to the buffer model for training and testing, but I kept having format incompatibility issues. Is ModelNet40 suitable for this model? Thank you so much!

在windows11下可以复现吗

Pretrained models and train dataset of 3DMatch

I am interested in your excellent work, and I want to know 1) what are the differences between your processed training dataset of 3DMatch and the training dataset provided by PREDATOR; 2) could you please provide the pre-trained models on 3DMatch to reproduce the results?

Code for entire module training

I'm trying to reproduce the result on your paper with 3DMatch dataset, however, I only could find the code for pre-training each module (e.g. "Ref", "Desc", "KeyPt", and "Inlier").
Could you tell me how to train all four modules simultaneously as the final step?

The pretrained model.

Thanks for sharing your codes. Could you please upload the pretrained model on KITTI for test. Thank you.

Possibility for MIT license?

Hi, great work! I'm wondering if you have considered having an open license such as MIT? Currently the license seems to be non-commercial.

why are here 2 decoder branch?

class EFCNN(Decoder)
class DetNet(Decoder)
Do the decoder branches in these two classes have different divisions of labor?

batch_size

Thanks for sharing your codes. When training the model, it seems that it does not support batch_size greater than 1. Will batch_size affect the accuracy of the model? Thank you.

multi-GPU training

Can it support multi-GPU training?