tidy up of tf-lift and lift => how to train test, or run pretrained model

[essie-chiang]

Hi, I could not run tf-lift no matter I will train or run pretrain, I followed lots of pre-issues:

1. Train with raw data is not supported?

   • Need read code to wrap raw dataset to h5 file? https://github.com/cvlab-epfl/tf-lift/issues/9 linke mentioned: You can look into datasets/eccv2016.py
   • Unable to open file scales-histogram-minsc-2.0.h5 https://github.com/cvlab-epfl/tf-lift/issues/5 Need read code and wrap data to scales-histogram-minsc-2.0.h5? But that is just training, could I train without this file?

2. Test:Need find LIFT repo However I find LIFT not able to run for it needs old Theano and old cuda...

My question is I am new to this area, and now is 2019, not reasonable to require support of a 2016 paper code. Could you just give some advise are there new projects that worth to follow? I am finding work that like LIFT, find better key point and image match than OpenCV.

Great thanks!

[kmyi]

Hi,

1. Train with raw data is not supported?

Indeed we cannot provide support for this. You'll need to study the code more carefully. To be honest, it's been a long time since this code was written, so we'll have to do the same.

2. Test:Need find LIFT repo

The test part should run without problems, given that you use the correct TF version, and the appropriate CuDNN libraries. I suggest using a docker instance with the correct version of TF.