Structural deformable shape models

Author: Steven Bergner

This is C++ research code implementing the image analysis system described in the thesis "Structural deformable models for robust object recognition".

The code is available on GitHub. Also see the project web page with a link to the thesis describing the ideas behind the implementation or have a look at the Overview of classes as an entry point to browse the code.

Data processing, multi-threaded optimization, and animated display of evolving models via GPU rendering are implemented from the ground up based on linear algebra from Eigen3, FFTW3, pthreads, OpenGL/freeglut, and Fox toolkit. Since there is no dependency on any external image processing code, the code may be fun to look at for learning about numerical software in C++, such as convolution with large filters via FFT (SmoothIntensitySensor::calcAllValues), Hessian determinant corner detection (CornerSensor::calcValue), PCA via SVD, spring-mass model solver, and many more components.

Brief description of classes

Node - a node in the geometry; has position, mass, velocitiy, attached forces, list of adjacent edges; has a Sensor attached

Edge - connects two nodes; has parameters for spring model, such as rest length, spring constant, damping constant

Model - holds information about geometry (Node, Edge), sensors, performs spring model calculation, and supplies all necessary state information to the controlling Brain

Dataset - handling the data set, its loading, display, and access by the Sensors; derived from class Sensor

Sensor - virtual class for accessing a source (data set or other sensor) and filtering this input; can preprocess a data sheet for the entire data set; base class for IntensitySensor, SmoothIntensitySensor, GradMagSensor, CornerSensor, etc.

Brain - controls a number of models and their states, manages a separate thread for spring model calculation and decision making

Searcher - performs a search for a given Model

ExpectationMap - list of EMDistribution objects that are generating specially distributed samples into the search space Model

MStructure - a node of the model structure containing a shape Model and links to other nodes, additionally it controls a Searcher by providing an ExpectationMap

StructTable - manages the complete structure graph, interpretations are used to find optimal paths selecting winners and appropriate spatial relations

SpeciesDB - is managing database queries selecting entries by given patterns, capabilities of STL template std::map are used to efficiently obtain query results