An differentiable pytorch implementation of SIFT patch descriptor

pytorch-sift

This is an differentiable pytorch implementation of SIFT patch descriptor. It is very slow for describing one patch, but quite fast for batch. It can be used for descriptop-based learning shape of affine feature.

UPD 08/2019 : pytorch-sift is added to kornia and available by kornia.features.SIFTDescriptor

There are different implementations of the SIFT on the web. I tried to match Michal Perdoch implementation, which gives high quality features for image retrieval CVPR2009. However, on planar datasets, it is inferior to vlfeat implementation. The main difference is gaussian weighting window parameters, so I have made a vlfeat-like version too. MP version weights patch center much more (see image below, left) and additionally crops everything outside the circular region. Right is vlfeat version


descriptor_mp_mode = SIFTNet(patch_size = 65,
                        sigma_type= 'hesamp',
                        masktype='CircularGauss')

descriptor_vlfeat_mode = SIFTNet(patch_size = 65,
                        sigma_type= 'vlfeat',
                        masktype='Gauss')

Results:

hpatches-results

OPENCV-SIFT - mAP 
   Easy     Hard      Tough     mean
-------  -------  ---------  -------
0.47788  0.20997  0.0967711  0.26154

VLFeat-SIFT - mAP 
    Easy      Hard      Tough      mean
--------  --------  ---------  --------
0.466584  0.203966  0.0935743  0.254708

PYTORCH-SIFT-VLFEAT-65 - mAP 
    Easy      Hard      Tough      mean
--------  --------  ---------  --------
0.472563  0.202458  0.0910371  0.255353

NUMPY-SIFT-VLFEAT-65 - mAP 
    Easy      Hard      Tough      mean
--------  --------  ---------  --------
0.449431  0.197918  0.0905395  0.245963

PYTORCH-SIFT-MP-65 - mAP 
    Easy      Hard      Tough      mean
--------  --------  ---------  --------
0.430887  0.184834  0.0832707  0.232997

NUMPY-SIFT-MP-65 - mAP 
    Easy     Hard      Tough      mean
--------  -------  ---------  --------
0.417296  0.18114  0.0820582  0.226832

Speed:

0.00246 s per 65x65 patch - numpy SIFT
0.00028 s per 65x65 patch - C++ SIFT
0.00074 s per 65x65 patch - CPU, 256 patches per batch
0.00038 s per 65x65 patch - GPU (GM940, mobile), 256 patches per batch
0.00038 s per 65x65 patch - GPU (GM940, mobile), 256 patches per batch

If you use this code for academic purposes, please cite the following paper:

@InProceedings{AffNet2018,
    title = {Repeatability Is Not Enough: Learning Affine Regions via Discriminability},
    author = {Dmytro Mishkin, Filip Radenovic, Jiri Matas},
    booktitle = {Proceedings of ECCV},
    year = 2018,
    month = sep
}