This is an keras implementation of SIFT patch descriptor. It is very slow for describing one patch, but quite fast for batch.

Here are comparisons to the original Michal Perdoch implementation. The benchmark is on W1BS dataset from WxBS: Wide Baseline Stereo Generalizations paper, figure 3. So there is no difference between versions in performance

Average performance on W1BS

Speed:

• 0.00246 s per 65x65 patch - numpy SIFT
• 0.00028 s per 65x65 patch - C++ SIFT
• 0.00100 s per 65x65 patch - CPU, 1 patch per batch
• 0.00087 s per 65x65 patch - CPU, 256 patches per batch
• 0.00236 s per 65x65 patch - GPU (GF940M), 1 patch per batch
• 0.00062 s per 65x65 patch - GPU (GF940M), 256 patches per batch

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

@article {tbd}