python-bp-disparity-maps

Computing Disparity-maps of images using loopy-belief-propogation algorithm

Requirements :

python-3.X
numpy
python-opencv-2

To run the program:

python main.py -imageno [IMAGE_NO] -maxdisp [MAX_DISP_ALLOWED] -smoothing [SMOOTHENING_CONSTANT]

Program-Specifications

First the given images are loaded & converted to grayscale, then VariableNodes and FactorNodes are initialized. Since the UnaryNode always returns the same message at each iteration it has been stored for efficiency.
Then we perform iterations, in each function Propogate is called which basically calls the buildBelief function of all VariableNodes. It inturn uses the precomputed unarynode message, and receives message from PairWiseFactorNode objects it is associated with.
Then message from Unary Node is a numpy array of size max disparity, and each element corresponds to the pixel magnitude difference at offset corresponding to that index
The message from PairWiseFactorNode is equal to a numpy array with all ones except at disparity of other associated VariableNode, where it is 0 and this numpy array is scaled up by smoothing factor / lambda. i.e. =>

CostArray = np.ones(MAX_DISP) * smoothing
CostArray[OTHER_NODE_DISP] = 0