search

NolenChen / 3DStructurePoints

77 stars 15 forks source link

Pre-train model for ShapeNet dataset #9

Open jimmy15923 opened 3 years ago

jimmy15923 commented 3 years ago

Hi! Thanks for sharing the amazing work. Would you share the pre-trained model on ShapeNet datset, which you mention in Sec 5.3 of the paper? And how do we perform testing on the un-align object? Thank you!

NolenChen commented 3 years ago

I've updated the pre-trained model here , and you can find real-scan chairs used in the paper in the folder /demo_data. I also updated the script test/test_structure_points.py for testing on un-aligned objects. You can perform testing on un-aligned objects by setting test_on_aligned to False, thus PCA will be used to reduce the rotation space.

jimmy15923 commented 3 years ago

Hi! Thanks so much for sharing your model and code:) Just a few more questions, how can I get the per-point features of structure points after alignment. I noticed that you apply a inv_rot_mats to transform the points, but the shape can't match when applying the to the inv_rot_mats raw per-point features. Also, do we need to apply farthest sampling on the test point cloud?

Thank you!

NolenChen commented 3 years ago

Hi! Thanks so much for sharing your model and code:) Just a few more questions, how can I get the per-point features of structure points after alignment. I noticed that you apply a inv_rot_mats to transform the points, but the shape can't match when applying the to the inv_rot_mats raw per-point features. Also, do we need to apply farthest sampling on the test point cloud?

Thank you!

You can use 'structure_points, features = model(batch_pts, return_weighted_feature=True)' to get the per-point features. But I do not quite understand the meaning of 'can't match when applying the to the inv_rot_mats raw per-point features'. Since the dimension of 'per-point features' is high, but the 'inv_rot_mats' can only be used to rotate 3-dimensional points. Can you describe more details about your scenario? As to the 'farthest sampling', based on my observation, it doesn't affect the performance a lot in our task. You may have a try, since you're dealing with real-scan objects, farthest sampling may help reduce the gap between real-scan objects and shapenet objects.

jimmy15923 commented 3 years ago

Hi! Sorry for the confusion. What I want are the features of the aligned structure points, e.g. (batch, 640, 16) for 16 structure points. Since you applied the inv_rot_mats on the original structure points, I thought it also needs to apply to the features of structure points?