Closed aprath1 closed 1 year ago
Hi @aprath1, Thanks for the question. Please see this issue from the Neural Descriptor Fields repo: https://github.com/anthonysimeonov/ndf_robot/issues/10. This portion of the codebase is very similar to ours. I believe you can voxelize the meshes and sample within the bounding box of the objects. Then use importance sampling to balance the number of positive and negative samples.
Ok, Thank you for your response, will try it out.
Could you please share how you are generating occupancy ground truths for the point cloud data generated with the 'shapenet_pcd_gen.py' ? The file only generates point cloud using pybullet simulation and I was wondering for these point clouds, how you have generated the occupancy ground truths. Is it possible to integrate it in the same (shapenet_pcd_gen) pipeline?
Thanks