search

zju3dv / GIFT

Code for "GIFT: Learning Transformation-Invariant Dense Visual Descriptors via Group CNNs" NeurIPS 2019
GNU Affero General Public License v3.0
191 stars 17 forks source link

Question about finetuning using GL3D #19

Open zhaotingy opened 2 years ago

zhaotingy commented 2 years ago

Great work!! I have one question in the finetuning stage using gl3d data. Do you apply any rotation or scale augmentation to the pair of images during the training? Please correct me if wrong. Looking at the code seems like no augmentation. If that's the case, how does it benefit from having the embedder as images in GL3D seems to have nearly 0 rotation. Thanks

liuyuan-pal commented 2 years ago

Thanks for your interest!

Since the embedder is implemented by conv layers on the rotation group, it can learn additional patterns defined on the group. This actually does not require augmentation of rotation and scaling.

When the rotation or scale differences are not very large, GIFT indeed has no obvious advantage over other descriptors.