3d_cascaded_seg

Demo: Super-Patch Algorithm

1. Run set_up_3dseg.m to set-up the path and compile the mex function.
2. Load example-model.mat in example/, which contains a model of human.
3. In super_patch, run seginfo = sp_main(face, vertex, sp_num, verbose); in which sp_num are number of super-patches and verbose=1 if you would like to manage the time.
4. To visualize the result, run: h = plot_mesh_segmentation(vertex,face,seginfo);
5. To see the training steps of this algorithm, please refer to sp_train.m

Demo: Cascaded Segmentation

1. load a model in meshsegBenchmark-1.0/data/seg/off/cascade_init
2. load learned_data.mat
3. tau = 0.5; verbose = 1;
4. run [cluster_matrix seginfo_matrix]= ... cascade_main(vertex, face, segstruct, seginfo, theta_matrix, mu_matrix, sigma_matrix, tau, verbose);
5. for i = 1:30 h = plot_mesh_segmentation(vertex,face,seginfo_matrix{i}) end

Directories:

• /Descriptors: Various discriptors for 3d models: HKS, ShapeDNA, etc.
• /example: example files to test the algorithm
• /features: features for 3d models
• /helper: some helper function
• /others: other functions for regular use.
• /visualization: functions to plot the segmentation results
• /super_patch: algorithms for super_patch calculation.
• /cascade: functions to train/run the cascade agglomeration algorithm

Explanation of the data structures:

• We adapted a matlab toolbox called "toolbox_graph" in this project, which provides an convenient way to load/operate on 3d models in off format.

• After reading the model using [vertex,face] = read_off(filename);

• 'vertex' is a 'n_vertice x 3' array specifying the coodinates of the vertices.
• 'face' is a 'n_face x 3' array, where face(:, i) specifies the vertice number of face i

• segmentations are stored in an one dimentional array, such as seginfo in the demo of super-patch algorithm, and seginfo(i) is the segmentation label of face i. If seginfo(i)==seginfo(j), it means that they are from the same segment.