This is a pipeline for aligning 3D point clouds and generate LSFM models. For more details, please check Optimal Step Nonrigid ICP Algorithms for Surface Registration.

how to load mesh clouds and transform them into the same size

how to make it run 10x times faster

what is the core algorithm to align mesh clouds from different point-of-view

how we build up the whole pipeline

Use menpo. For details, please check Menpo.

p = Pipeline(base_model_path)
lsfm, logs = p.run(input_path)

// to use LSFM model, use lsfm
// to have a look at training logs, take logs

LSFM will run for many epochs. Each epoch will run for many iterations. Weights for each epoch and other hyper-parameters are already handled in config.ini.

The config.ini is as follows:

[DEFAULT]

• DEFAULT_OUTPUT_PATH = ./output save models and intermediate data into that directory.
• DEFAULT_STIFFNESS_WEIGHTS = [50, 20, 5, 2, 0.8, 0.5, 0.35, 0.2] stiffness weights for each epoch
• VAR = [85, 300, 220] mesh cloud variance in 3 dimensions
• CENTER = [0, 0, 0] mesh cloud center
• SOLVER = umfpack mathematical solver, 'umfpack' or 'naive'
• MAX_ITER = 10 max number of iterations for each epoch
• EPS = 1e-3 epsilon
• MAX_NUM_POINTS = 2000 max number of points retained for each mesh cloud
• N_COMPONENTS = 0.997 number of components (mesh clouds) retained, if it's within $(0,1)$, then k largest components are retained to reach the variance of N_COMPONENTS; otherwise N_COMPONENTS largest components are remained.
• MESH_FILE_EXTENSIONS = [OBJ,] which kind of files are mesh files containing point clouds to be aligned
• SAVING_FREQUENCY = 50 for every SAVING_FREQUENCY mesh files processed, we should save the current pipeline into output_path as backup
• IS_PREEMPTIVE = True whether to resume from previously results, if true, try to load the most updated pipeline model saved from output_path and re-run the pipeline from that point
• VERBOSE = True

Also, you can customize the above parameters and use them to construct your Pipeline model with the construtor function. In particular, a constructor parameter data_weights should be of the same length as that of stiffness_weights. In general, you can take it as None and ignore it.

For the details about implementing and testing different parts of our pipeline, please check LSFM notebook.