search

vcg-uvic / lf-net-release

Code Release for LF-Net: Learning Local Features from Images
Other
314 stars 66 forks source link

Test on Roxford5k benchmark #8

Open wickai opened 5 years ago

wickai commented 5 years ago

I am doing some research on image retrieval. I find your method really innovative. But when I use the model you gave to evaluate on Roxford5k benchmark (Medium level) extracting 1000 keypoints, the conclusion indicators are very low, mAP only 26.7, mp@10 50.14. I also did RANSAC after extracting keypoints.

Is this a reasonable conclusion or I did somethin wrong? Easier queries seems right, but a little difficult queries with view point change turn out wrong.

image

thanks~

kmyi commented 5 years ago

lfnet is trained for stereo matching, so we honestly have little idea on how it would perform for image retrieval. A few things that we've noticed is that ratio tests or other heuristics involved in traditional matching could be harmful instead. Thus, a bit of care must be taken when plugging the features into a sophisticated pipeline. Which is true for all local features I guess...