This software uses a single Adaboost classifier for trainig purposes. The cascade-classifieris yet to be implemented.
- numpy
- scipy
- Pillow (for drawing boxes)
- joblib
-
For 2D object detection, face data from MIT was used.The data is stored as it is faces_data/ . Images are in .pgm & other formats - only .pgm files are read.
-
For 3D object detection, a synthetic dataset was composed of spheres & non-spheres( cubes + cylinders). Both train & test contain 10 spheres & 10 non-spheres respectively. Each file consists of 151515 voxels, with cubes/spheres/cylinders being centred at (7,7,7) but with parameters for each shape being chosen at random.It was generated using this script
2D Object detection was based on the famous Viola-Jones paper.
3D object detection based on Haarlike features was used by Wesrag et al for organ detection, as part of the MICCAI Grand Challenge: Prostate MR Image Segmentation 2012. The Haarfeatures include the same fetures mentioned in the [paper](/resources/A Generic Approach to Organ Detection Using 3D Haar-Like Features .pdf)
The object detection scripts can be executed from root of the repository
For running 2D object detection
$: python 2D_HaarlikeFeatures/ViolaJones.py
For running 3D object detection
$:python 3D_HaarlikeFeatures/threeD_violajones.py
For 3D object detection, the best results that could be attained
For 2D object detection the faces could only be detected to a certain extent. More optimisation needs to be done. original image