Multi-GPU Implementations && Keypoints Generation

[qsisi]

Hello! I have two questions:

1. As far as I can see, the current released version doesn't support Multi-GPU implementation, the hyper-params are set clearly for Multi-GPU (train_batch_size is set to 76, I manage to run the code in my single TITAN V with 11 GB memory by setting it from 76 -> 16) , so could you release the Multi-GPU implementation version?

2. Could you tell me what is the strategy of selecting 5k keypoints in 3DMatch?

Best.

[aosheng1996]

Hi, @qsisi, thanks for your interest in our work!

1. You only need to modify these two lines of code to run on multiple GPUs: https://github.com/QingyongHu/SpinNet/blob/5581e7d184bc3b4d525d5b5e58777ea04dfdc9ab/ThreeDMatch/Train/train.py#L3 https://github.com/QingyongHu/SpinNet/blob/5581e7d184bc3b4d525d5b5e58777ea04dfdc9ab/ThreeDMatch/Train/trainer.py#L34

2. We randomly sample 5000 points as the keypoints to evaluate feature descriptors.

Best, Sheng

[qsisi]

Thanks very much for the timely reply which solves the aforementioned questions.

Also, may I ask you how to generate the *.pkl files in the ./data/3DMatch/patches/train that contain the anchors and positives? I notice that there is a fuse_fragments_3DMatch.py file, but it seems like it functions as fusing depth images in 3DMatch into *.ply fragments, but not generate *.pkl file from *.ply fragments. So can you release the data pre-processing code which converts *.ply to *.pkl ? That would be much helpful for further understand this project.

Best.