libHOG
libHOG copied to clipboard
forresti
fast and energy-efficient computation of HOG features
libHOG
About
libHOG is a library that computes Histogram of Oriented Gradient feature pyramids. The benefit of libHOG over other HOG implementations is speed. On an Intel i7-3930k 6-core CPU, we measured the following results:
| Frame Rate | Speedup | Energy per frame (Joules) | |
|---|---|---|---|
| voc-release5 | 2.44 fps | 1.0x | 57.4 J/frame |
| Dollar | 5.88 fps | 2.4x | 26.4 J/frame |
| FFLD-serial | 4.59 fps | 1.9x | 29.9 J/frame |
| FFLD-OpenMP | 19.6 fps | 8.0x | 9.44 J/frame |
| libHOG (this codebase) | 58.8 fps | 24.1x | 3.15 J/frame |
The output of libHOG is numerically equivalent to the HOG features in the Deformable Model Parts model (voc-release5) codebase, with a 24x speedup. For further speedups (over 70fps), we also offer the option to use L1-norm (instead of the traditional L2-norm) when calculating the gradients. If you find libHOG useful, please consider citing the libHOG paper:
@inproceedings{libHOG,
Author = {Forrest N. Iandola and Matthew W. Moskewicz and Kurt Keutzer},
Title = {libHOG: Energy-Efficient Histogram of Oriented Gradient Computation},
Booktitle = {ITSC},
Year = {2015}
}
Getting Started
- install OpenCV
- point Makefile to OpenCV
- mkdir tests_output; mkdir build;
- make
basic example on how to use libHOG in C++:
//computing HOG features on my_image.jpg
cv::Mat img = cv::imread("my_image.jpg");
libHOG lHOG();
lHOG.compute_pyramid(img);
//HOG pyramid is located here: lHOG.hogBuffer_blocks[scale]
//pyramid dimensions: lHOG.hogHeight[scale], lHOG.hogWidth[scale]
//number of scales: lHOG.hogHeight.size()
see src/libHOG_cli.cpp for an additional example of how to use libHOG
linking your project with libHOG: include libHOG.h, and dynamically link with libHOG.so
forresti