CPU+GPU Dense CRF

Header-only Dense Conditional Random Field (CRF) in C++ and CUDA with easy-to-use interface for 2D image and unorganized data (such as point cloud).

NOTE: This version does not support gradient computation / back propagation of parameters.

For the above example with size $320\times 240$ with 21 classes and 10 iterations, the CPU version takes 225ms while GPU version takes 88ms (statistics computed using Linux time utility).

Compile & Run

This is a header only library so all you need to do is to copy all the header files in include/ directory into your project.

For compiling the examples, you can use the cmake build system:

mkdir build && cd build
cmake ..
make example_cpu
./example_cpu ../examples/im1.ppm ../examples/anno1.ppm ../examples/res1_cpu.ppm
make example_gpu
./example_gpu ../examples/im1.ppm ../examples/anno1.ppm ../examples/res1_gpu.ppm

Usage

CPU Version

#include "densecrf_cpu.h"
#include "pairwise_cpu.h"
using namespace DenseCRF;

// Setup the CRF model
DenseCRFCPU<M> crf(W * H);
crf.setUnaryEnergyFromLabel( label, 0.5 );
// add a color independent term (feature = pixel location 0..W-1, 0..H-1)
// x_stddev = 3
// y_stddev = 3
// weight = 3
auto* smoothnessPairwise = PottsPotentialCPU<M, 2>::FromImage<>(W, H, 3.0, 3.0);
crf.addPairwiseEnergy( smoothnessPairwise );
// add a color dependent term (feature = xyrgb)
// x_stddev = 60
// y_stddev = 60
// r_stddev = g_stddev = b_stddev = 20
// weight = 10
auto* appearancePairwise = PottsPotentialCPU<M, 5>::FromImage<unsigned char>(W, H, 10.0, 60.0, im, 20.0);
crf.addPairwiseEnergy( appearancePairwise );
// Do map inference
crf.inference(10, true);
short * map = crf.getMap();

GPU Version

Please note all pointers should be CUDA device pointer.

#include "densecrf_gpu.cuh"
#include "pairwise_gpu.cuh"
using namespace DenseCRF;

// Setup the CRF model
DenseCRFGPU<M> crf(W * H);
crf.setUnaryEnergyFromLabel( labelGPU, 0.5 );
// add a color independent term (feature = pixel location 0..W-1, 0..H-1)
// x_stddev = 3
// y_stddev = 3
// weight = 3
auto* smoothnessPairwise = PottsPotentialGPU<M, 2>::FromImage<>(W, H, 3.0, 3.0);
crf.addPairwiseEnergy( smoothnessPairwise );
// add a color dependent term (feature = xyrgb)
// x_stddev = 60
// y_stddev = 60
// r_stddev = g_stddev = b_stddev = 20
// weight = 10
auto* appearancePairwise = PottsPotentialGPU<M, 5>::FromImage<float>(W, H, 10.0, 60.0, rgbFeatGPU, 20.0);
crf.addPairwiseEnergy( appearancePairwise );
// Do map inference
crf.inference(10, true);
short * mapGPU = crf.getMap();

References

Krähenbühl, Philipp, and Vladlen Koltun. "Efficient inference in fully connected crfs with gaussian edge potentials." Advances in neural information processing systems. 2011.

Adams, Andrew, Jongmin Baek, and Myers Abraham Davis. "Fast high‐dimensional filtering using the permutohedral lattice." Computer Graphics Forum. Vol. 29. No. 2. Oxford, UK: Blackwell Publishing Ltd, 2010.

heiwang1997 / DenseCRF