"K-Means"

K-means image segmentation with CUDA (a "to be optimized" single threaded kernel implementation).

Homework assignment

Students are asked to parallelize a naive implementation of the K-Means image segmentation algorithm. A single threaded kernel is provided in k_means_cuda.cu and corresponds to the code that needs to be optimized. Everything else can be seen as a boiler plate code.

Requirements

• a CUDA capable device with compute capability >= 2.0
• the appropriate CUDA-toolkit needs to be installed

Compilation

# run from repo dir
nvcc main.cpp external/Color.cpp external/utils.cpp k_means_cuda.cu -o out/k_means_cuda
# run ./out/k_means_cuda

When running out/kmeans_cuda

• the user is asked to pick an image to be segmented
• for quick tests, it is recommanded to pick the small one by tapping s
• for measuring the gained performance of your optimizations it is recommanded to pick the big one by tapping l
• input & ouptup images are loaded & saved to ./data dir
• increase k to stress your approache

TODOs

• upload the homework assignement's PDF
• upload ./data/input_large.bmp
• translate REAMDE.md to french

Attribution

This code uses Bitmap Image Reader Writer Library by @ArashPartow.