This short project intended to evaluate the quality of appearance models. It did so by synthesising thousands of examples from the model and calculating their distances w.r.t. to the model they originated from. The implementation was in C++ and MATLAB. The MATLAB component is public and fully open.
Description of the framework
Step 1: Register data set
Discrepancy image of a pair as registration proceeds
Step 2: Use registration output to build an appearance model
Models are poor and the start, but later sharpen up
Step 3: Synthesise from the model
Obtain thousands of images like the above
Step 4: Pre-process syntheses
Crop the unwanted artefacts if any exist and stretch all images to fit the entire space nicely.
Step 5: Calculate shuffle distance
Use the shuffle transform to derive distance for each possible pairing between the model training set and syntheses.
There are several related publications in the research workspace and the publications page.