real-stanford / garmentnets

[ICCV 2021] GarmentNets: Category-Level Pose Estimation for Garments via Canonical Space Shape Completion
https://garmentnets.cs.columbia.edu/
59 stars 8 forks source link

Question regarding the Correspondence Distance #6

Closed leongkhing closed 7 months ago

leongkhing commented 7 months ago

Hi! Thank you for the interesting work! I have a question regarding the Correspondence Distance metric used in the paper. If I understand correctly, does the Correspondence Distance measure the L2 distance between the predicted point and its corresponding ground truth point in NOCS?

Also in the evaluation, does the nocs_pc_error_distance correspond to the Correspondence Distance? If so, may I know if it is normal only to get 0.3830191791057586 for the nocs_pc_error_distance using the sample datasets? I assume the unit is in meters? Since the chamfer_symmetrical_nocs is 0.01829747862776, which matches the result in the paper. Could the high nocs_pc_error_distance be because of the evaluation is only done on the sample dataset?

Please let me know if I understand it correctly. Thank you in advance!