DEEPSDF.md

Training DeepSDF

Dataset

To shape the train embedding model a json file is expected the provides relative paths to each shape you want to embed. The json file should follow this format:

[
    {
        'path': relative path to first npz
    },
    ...
    {
        'path': relative path to last npz
    }
]

Each entry should specify an npz that stores the SDF samples.

The convert.py file has a util function to create such a json for you from a folder of npz files:

convert.create_sample_json(sdf_folder)

However, you should only train a shape embedder on shapes that will not be in the validation or test set of the later pose estimation step, so you likely want to create the json file manually

Training

IM_ROOT should point to the folder of SDF samples. SAVE_PATH should point to where you want the checkpoints saved. config.yml holds the network and training configurations that seem to work well

python embed_shape.py --im_root IM_ROOT --yaml_file configs/embed_config.yml --save_path SAVE_PATH --file_list JSON_FILE

Verification

After training the shape embedding model, you can check how well the model visually captures the distribution of shapes:

infer_shape.infer_mean_mesh(MODEL_CKPT, config.yml, SAVE_LOC, SDF_RESOLUTION)

The above function will generate a uniformly sampled SDF from the mean latent vector and then perform marching cubes to obtain a mesh. The resulting file can be loaded into a mesh viewwer (MITKWorkbench, pyrender). You should see a reasonable organ shape.

Individual shapes from the training set can also be visualized, where LATENT_IDX should specify the index in the training json of the shape you want ot visualize:

infer_shape.infer_mesh(MODEL_CKPT, config.yml, LATENT_IDX, SAVE_LOC, SDF_RESOLUTION)