Goals

Main Questions to Address

Determine the best data sources to use for the obstructed image database

 1. COIL100 dataset is a collection of successive 5 degree shots for a total of 72 shots of an object, with 100 objects documented
    ("Real-time 100 object recognition system" IEEE conference on robotics and Automation 1996)

Determine if artificial or realistic obstructions are better suited for study

 1. Artificial obstructions are easily repeatable for research purposes
 2. Realistic camera grain and object-over-object obstructions with many gaps provide real world usage benefits if addressed

Determine the best data set to implement when training NNs (Neural Networks)

 1. Image pixel data per feature extraction segment from kmeans
 2. Centroid per feature extraction segment
 3. Chain codes from detatched edges
 4. Delaunay graphs & triangulation
      1. Angles connecting the centroids from the triangulation
 5. 3D Matrix of centroid graph for subsequent matching processor augmentations

Subsequent Ideas To Improve On From Original Obstruction Codebase In Java

 1. Weka Machine Learning algorithms have been thoroughly explored; Begin with Deep Learning Neural Networks instead
 2. In order to more wholly complete an image, incorperate a few new preprocessing methods to further enhance the NN understanding of the whole image
      1. Implement a basic shape recognition algorithm
      2. Implement a generative edge completion algorithm to recognize common edges that are actually relevant to each other to combine into a kmeans region
 3. Implement intrinsics datatypes on all matrices and matching algorithms to allow faster and more accurate results to be produced
 4. Improve weight calculations when determining match results with a Local-Global Weight Calculation
      1. Cluster exemplar images in the database together with labels based on the object they represent
      2. When matching a sample against the database, score each image match locally
      3. Average the local scores per relation-cluster to determine a general probability of an image match
      4. Weight the previous average cluster score globally with any local score in its cluster over or under a certain threshold
 5. Improve matching time significantly in cooperation with intrinsics using an algorithm similar to cinnebench image render test