Third-Order Edge Detection

This is a OpenMP C++ and CUDA GPU implementation of Third-Order Edge Detection (TOED). See the referenced paper for more information. The matlab code of the paper can be found in Yuliang's github page.

@article{kimia2018differential,
  title={Differential geometry in edge detection: accurate estimation of position, orientation and curvature},
  author={Kimia, Benjamin B and Li, Xiaoyan and Guo, Yuliang and Tamrakar, Amir},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  volume={41},
  number={7},
  pages={1573--1586},
  year={2018},
  publisher={IEEE}
}

Dependencies

The code has been tested in Linux-based system with the following versions of dependencies:
(1) cuda/11.1.1 or higher
(2) (Optional) OpenCV 3.X or above (only used to read an image and access image pixel values)
Note that:

• CUDA version depends on the GPU. Please have it checked to use the correct cuda version, e.g., using the $ nvidia-smi command.
  
• If you do not use OpenCV, set OPENCV_SUPPORT in indices.hpp to false, and command out the include paths and library paths in the makefiles.

🛠️ Setup and Run the code

There are two make files to build and compile the code: (i) makefile.gpu_cpu works for all the files, including the GPU and the CPU code. (ii) makefile.cpu works only for the CPU code with double precision. You can do either one of them to build and compile the code by,

$ make -f makefile.gpu_cpu  // or make -f makefile.cpu (for CPU-only version)

Make sure you have changed the paths for CUDA, OpenCV, etc. in the makefiles. Once everything works perfectly, proceed to execute the code by

• For GPU+CPU

$ ./TOED <name_of_input_image> <number of CPU threads> <gpu id>

• For CPU only:

$ ./TOED <name_of_input_image> <number of CPU threads>

The argument <name_of_input_image> is mandatory while the rest are optional. If OpenCV is supported, any type of images should be supported. Otherwise, only .pgm image file is accepted. A few sample images are provided in ./input_images/, so you can, for example, run the code using:

$ ./TOED ./input_images/euroc_sample_img.png 4

You can clear out all the *.o files by

$ make -f makefile.gpu_cpu clean

📺 Display edges and orientations

After a successful run, lists of subpixel edges are written in text files named data_final_output_cpu.txt under ./output_files/. You can use the matlab file in ./draw_edges_by_matlab/draw_edges_from_list.m to plot the edges of the input image, or ./draw_edges_by_matlab/draw_edges_orient_from_list.m to plot the edges and their orientations of the input image.

⏲️ Some test results (To be updated)

CPU: Intel(R) Xeon(R) Gold 6242 CPU @ 2.80GHz
GPU: NVIDIA QuadroRTX 6000
Double Precision Test:
==> CPU Test (OpenMP 4 threads)
============================================

• Time of image convolution (OpenMP): 261.332 (ms)
  
• Time of NMS (OpenMP): 2.62396 (ms)
  
  

==> GPU Test
=============================================

• GPU Convolution time = 16.5035 ms
  
• GPU NMS time = 0.2988 ms
  
  

Single Precision Test:
==> CPU Test (OpenMP 4 threads)
=============================================

• Time of image convolution (OpenMP): 257.492 (ms)
  
• Time of NMS (OpenMP): 2.32577 (ms)
  
  

==> GPU Test
=============================================

• GPU Convolution time = 0.7848 ms
  
• GPU NMS time = 0.2951 ms
  

Note that the above timings could change according to the input image size, the power of CPU/GPU.

⤴️ ChangeLogs

Jul. 23, 2023: The curvel formation code is also included in thie repo for completeness. It only accepts double precision third-order edges for now. Single precision will be made available for curvel formation in the near future. Feb. 4, 2025: Fix a few bugs and also support OpenCV to read an image