Insanely fast CUDA LATCH 512-bit binary descriptor for computer vision (upright)
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Fastest implementation of an UPRIGHT (no rotation) LATCH 512-bit binary feature descriptor as described in the 2015 paper by Levi and Hassner:

"LATCH: Learned Arrangements of Three Patch Codes"

See also the ECCV 2016 Descriptor Workshop paper, of which I am a coauthor:

"The CUDA LATCH Binary Descriptor"

And the original LATCH project's website:

Note once again that this is an UPRIGHT LATCH, a.k.a. ULATCH. A fast rotation- and scale-invariant version is also available on my GitHub.

This implementation is insanely fast, matching or beating the much simpler ORB descriptor despite outputting twice as many bits AND being a superior descriptor.

A key insight responsible for much of the performance of this laboriously crafted CUDA kernel is due to Christopher Parker ( to whom I am extremely grateful.

CUDA CC 3.0 or higher is required.

All functionality is contained in the files UCLATCH.h and 'main.cpp' is simply a sample test harness with example usage and performance testing.