AAM Property Files

AAM Property (*.aamprops) files describe how to build an AAM, including various parameters necessary for the builder.
They are used by the build_aam tool.
Here's an example:

// Define type of AAM (basic_aam,shape_aam,comp_aam,direct_aam,raw_regression_aam,grad_regression_aam) aam_type: basic_aam // Define model directory and base name of APM and AAM files model_dir: ./ am_name: isbe_comp aam_name: isbe_comp // Parameter displacements used during search displacement_x: 2 displacement_y: 2 max_disp: 0.125 scaledisplacement_x: 0.05 scaledisplacement_y: 0.05 tex_displacement: 0.05 displacement_params: 0.5 n_per_eg: 2 // Maximum number of images to sample from (overrides that supplied in SMD) max_n_samples: 10 rob_kernel: None

The types of AAM available include:

basic_aam : Plain AAM, trained by estimating gradient d(residual)/d(param), as descibed in PAMI paper
raw_regression_aam: Plain AAM but trained using regression (as described in early papers)
grad_regression_aam: Plain AAM but trained using regression to estimate gradient (as described in BMVC02 paper)
shape_aam: AAM in which linear shape update predicted from current texture error
comp_aam: AAM in which shape update predicted from current texture error and applied using a compositional approach
direct_aam: AAM in which shape predicted from current texture parameters

Recent experiments on faces (described in BMVC02 paper and related report) suggest that the grad_regression_aam is probably the best default option, though the basic_aam performs well and is quicker to train.
Surprisingly the shape_aam and comp_aam seem to match texture slightly better, though it is somewhat problem dependent.