How to solve CUDA out of memory

[zlynpu]

Hello, after seeing your article, I would like to ask you how to solve the problem of CUDA out of memory of Geotransformer on the scan2cad data set. I use NVIDIA 3090 and use voxel sampling(voxel size=2.5cm) to sample the scene points to within 30,000, but this problem still occurs. If possible, can you give me your code for reference?

[zhiyuanYU134]

Hello, after seeing your article, I would like to ask you how to solve the problem of CUDA out of memory of Geotransformer on the scan2cad data set. I use NVIDIA 3090 and use voxel sampling(voxel size=2.5cm) to sample the scene points to within 30,000, but this problem still occurs. If possible, can you give me your code for reference?

Due to the excessive number of superpoints, Geotransformer cannot be trained or tested on the Scan2cad dataset. Therefore, we adopted some methods to approximate the performance of Geotransformer. We limit the number of superpoints that interact with the anchor superpoint. This number will be set as large as possible to approximate the performance of Geotransformer. You may have to implement this code yourself.
Another possible way is to downsample again based on the original superpoint. The features of these sparser superpoints are obtained by aggregating the original superpoints. After these sparse superpoints learn geometric embeddings, interpolation is used to calculate the features of the original superpoints.

[zlynpu]

Thank you very much for your answer! I have another question about the Negative Log-likelihood Loss in the article. How do you solve the correspondences based on the transformation matrices of different instances? Because from my understanding, there is only one cad model but it corresponds to multiple instances in the scene.