BRIEF-Binary-Descriptor-AVX2

A fast BRIEF Implementation for 256-dimensional 64-bit Binary Descriptors

Description

This is a C++ method that allows you to calculate the BRIEF binary feature descriptors.

Performance

This method is approx. 8x faster than OpenCV's BRIEF method. The reason is that I am using a 64-bit strategy as opposed to the 8-bit strategy of OpenCV. Furthermore, my code is aggressively optimized. That is, the kernel contains mostly packed avx2 instructions.

Timing for a set of 500 features for 256-dimensional binary descriptors:

• OpenCV: 2.5 ms
• this version: 0.31 ms

Features

• 32-byte aligned plain arrays instead stl vectors
• Gaussian pattern is divided into 4 256x1 contiguous arrays instead of a 256x4 matrices. Allows vectorization.
• Complete unrolling of the i = 0 ... 256 loop
• AVX2 masking (_mm256_movemask_epi8) and comparing (_mm256_cmpeq_epi8) and load-n-store removes need of bitshifting

Photos

PHOTO1: Assembly output shows nice and consistent packed instructions

PHOTO2: Intel intrinsics handle gathering and ordering of bits in a packed way:

To-Do

• Unroll loops by pre-processor
• Try to speed up memory-bound bottleneck by AVX2 gather/scatter instructions

Dependencies

• m4 (macro pre-processor)
• Intel ispc SIMD compiler

How-To Install

Have a look at the CMake file. Just run it with your individual paths.

Contact

Feel free to contact me if you have questions or just want to chat about it.