Dataset creation with 3DMatch dataset for training

[surabhi96]

Hi @zgojcic

I am trying to train 3DSmoothNet with the 3DMatch dataset. I wanted to know how are the descriptors (dataset) generated for main_cnn.py? As in how are pairs of rgbd images chosen (from which the keypoint descriptors are computed)?

Thanks,
Surabhi

[zgojcic]

We first convert the RGBD images to point cloud fragments as done in 3DMatch dataset, we then choose the fragments with at least 30% overlap to sample the correspondences. You can check the code here in how we compute the overlap https://github.com/overlappredator/OverlapPredator/blob/main/scripts/cal_overlap.py

[zgojcic]

Closing due to inactivity. Please feel free to open a new issue if the answer above does not solve your problems.

Best
Zan