=================================================================================== NMS Benchmarking Framework

CUDA implementation of the algorithm described in the paper:

"Work-Efficient Parallel Non-Maximum Suppression Kernels"

http://dx.doi.org/10.1093/comjnl/bxaa108

The Computer Journal

David Oro, Carles Fernández, Xavier Martorell, Javier Hernando

• Requirements:

       1. GCC Compiler v5.0 or greater
       2. CUDA Toolkit v6.0 or greater
       3. NVIDIA GPU with Compute Capability 3.2 or greater
  

• Build instructions:

       1. Set the GPU_ARCH and SM_ARCH variables in the Makefile according to 
          the underlying NVIDIA GPU architecture of your computer. For further 
          details, please refer to our GitHub Wiki page:
  
                https://github.com/hertasecurity/gpu-nms/wiki
  
       2. Set your CUDA installation path in the Makefile (CUDA_HEADERS and 
          CUDA_LIBS variables)
  
       3. Compile the source code:   make
  

• Execution:

       * You can run the GPU NMS benchmark using a comma-separated input file 
         containing the list of detected objects in the following format:
  
                xcoordinate,ycoordinate,width,score
  
       * We provide a sample input file "detections.txt" obtained after having 
         executed a face detector over the "oscars.png" file.
  
       * The GPU NMS benchmark must be executed as follows:
  
                ./nmstest  detections.txt  output.txt
  
       * The application should then return the computation time of both the MAP 
         and REDUCE GPU NMS kernels and write the results in the "output.txt" file.
  
       * Finally, you can visualize both the input (pre-NMS) and the output 
         (post-NMS) with the "drawrectangles" Python script. For example:
  
                ./drawrectangles  detections.txt
  
         Or:
  
                ./drawrectangles  output.txt
  
         The graphical output is stored in the "oscarsdets.png" file
  

• IMPORTANT:

       * The source code must be compiled to the microarchitecture matching the 
         GPU platform during execution (check GPU_ARCH and SM_ARCH variables 
         in the Makefile).
  
       * If the NMS algorithm is not capable of properly merging the candidate 
         windows, re-check the GPU_ARCH and SM_ARCH variables and then 
         recompile the code.
  
       * This GPU NMS benchmark is limited to a maximum of 4096 detected 
         objects per input. If you want to increase this limit, please 
         modify the MAX_DETECTIONS constant in the "nms.cu" file.