How about the accuracy of GIFT at a larger scale ratio of image pair ?

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

How about the accuracy of GIFT at a larger scale ratio of image pair ? #7

Closed xiaoguang12 closed 4 years ago

xiaoguang12 commented 4 years ago

Hi thanks for your outstanding work, i wonder How about the accuracy of GIFT at a larger scale ratio of image pair such ac scale ratio of or even larger. in your paper you seem just test at 2^1.5. thanks looking forward to your reply.

Best wish xiaoguang

xiaoguang12 commented 4 years ago

scale ratio of 10 or even more

liuyuan-pal commented 4 years ago

We didn't do systematical experiments to test it under a large scale ratio like this, because it's relatively rare in matching natural images. Theoretically, if sampling more elements on the scale axis (currently only 5 elements), then the invariance of GIFT to scale will be stronger. But, meanwhile, sampling more elements means more computational complexity.

xiaoguang12 commented 4 years ago

Ok i see thanks for your reply.