PCL Developers blog

Roadmap

This is a detailed roadmap for my GSoC 2012 project. The boldface writings represent my current activities.

  • Module keypoints:
    • First approach to the object recognition research field: documentation on techniques at the state of the art.
    • Test functionalities for the new 3D object recognition tutorial and the existing detectors in PCL.
    • Implement a framework for simple keypoint detection evaluation:
      • Determine detectors’ limits and detectors’ saliency.
      • Structured comparison between the existing keypoint detectors in PCL:
        • Detectors under testing: Harris 2D, 3D, 6D, SIFT, NARF, Uniform sampling.
        • Compute absolute repeatability.
        • Compute relative repeatability.
      • Visualize results by means of the PCLVisualizer.
    • Extend the framework in order to perform tests on complete datasets:
      • Structured comparison between the existing keypoint detectors in PCL:
        • Detectors under testing: Harris 3D, 6D, SIFT, NARF, Uniform sampling.
        • Study of the detectors’ parameters.
        • Compute average absoulute repeatability on the overall dataset at issue.
        • Compute average relative repeatability on the overall dataset at issue.
        • Perform an effective analysis about detectors’ time performances:
          • Compute the extraction time for each keypoint detector.
        • Summarize the collected data by means of appropriate tables and graphs.
      • Tests execution on two different datasets:
        • A kinect-based dataset (Dataset 5 - Kinect).
          • Detectors: Harris 3D, 6D, SIFT, NARF, Uniform sampling.
        • A synthetic dataset (Dataset 2 - Stanford).
        • Detectors: Harris 3D, NARF, Uniform sampling.
        • Both the datasets can be found at http://vision.deis.unibo.it/SHOT/ .
    • Porting of the ISS detector in PCL.
      • Read the documentation paper.
      • Take some knowledge about the code that has already been implemented.
      • Design of the ISSKeypoint3D class:
        • Define the input parameters of the detector.
        • Define the output parameters of the detector.
        • Define the first pipeline operations needed to compute the interest points.
        • Implement the skeleton of the class.
        • Implement the methods needed to obtain a basic functionality of the detector.
      • Refinement of the ISSKeypoint3D class:
        • Take knowledge about the time performances of the detector by means of a basic evaluation framework.
        • Take knowledge about the repeatability performances of the detector by means of a basic evaluation framework.
        • Detect the bottleneck (if any).
        • Enhance the time performances of the detector by using appropriate data structures.
        • Enhance the time performances of the detector by using the OpenMP directives.
        • Add the boundary estimation if it is possible and it does not decrease the performances of the detector (both time performances and repeatability performances).
    • Test the performances of the ISS detector.
      • Use the skeleton of the evaluation frameworks developed at the beginning of the GSoC.
      • Test the repeatability performances.
      • Test the time performances.
      • Test the time performances related to the OpenMP optimization.
      • Compare the performances of ISS with those of the other dectors already tested.
    • Tests execution on two different datasets (as for the first part of the my GSoC roadmap).
  • Tools
    • Develop a new point type to handle monochrome images.
      • Name: pcl::Intensity.
      • Number of fields: 1 (type: uint8_t).
    • Develop a PNG to PCD converter.
      • Input:
        • Name of the PNG input file.
        • Name of the PNG output file.
      • Output:
        • A PCD file that represents the conversion of the PNG input file.