Tensorflow 2 implementation of the Convolutional Mesh Autoencoder (CoMA)

Re-implementation of the Convolutional Mesh Autoencoder (CoMA) proposed by Ranjan et al. in their paper Generating 3D faces using Convolutional Mesh Autoencoders.

Data

The data for training and evaluation the models is available on the project page of the original paper.

Training

The utility

./train_models.sh

can be used to train the models used for evaluation.

Evaluation

The evaluation utility is split, so that prediction and error calculation can be done on different computation instances.

To predict the test data using the trained models use:

./calculate_predictions.sh

To calculate the errors given the prediction output of the above utility, use:

./calculate_errors.sh

Main.py

The autoencoders main file can be invoked from commandline via the main.py:

python main.py --name run-name --data-folder /path/to/preprocessed/data --mode train|test|latent

Output of python main.py --help:

Convolutional Mesh Autoencoder written for Tensorflow 2

optional arguments:
  -h, --help            show this help message and exit
  --name NAME           The name of the run (used for checkpoints and tensorboard)
  --data-dir DATA_DIR   Path to the data folder containing train.npy and test.npy
  --batch-size BATCH_SIZE
                        The batch size to be used (default is 16)
  --num-epochs NUM_EPOCHS
                        The number of training epochs (default is 300)
  --initial-epoch INITIAL_EPOCH
                        The initial epoch, useful for continue training on an
                        existing run (default 0)
  --latent-vector-length LATENT_VECTOR_LENGTH
                        The size of the latent vector (default is 8)
  --validation-frequency VALIDATION_FREQUENCY
                        The validation frequency (default is 10)
  --learning-rate LEARNING_RATE
                        The learning rate (default is 8e-3)
  --random-seed RANDOM_SEED
                        The random seed (default is 8)
  --template-mesh TEMPLATE_MESH
                        Path to the template mesh (default is data/template.obj)
  --mode MODE           The mode to run in train|test|latent (default is train)
  --sanity-check SANITY_CHECK
                        Whether or not sanity check should be performed (default is
                        False)
  --coma-model-dir COMA_MODEL_DIR
                        The directory holding checkpoints and tensorboard, such as
                        coma-model/tensorboard or coma-model/checkpoint (default is
                        coma-model)
  --visualize-during-training VISUALIZE_DURING_TRAINING
                        Whether the meshes should be visualized in tensorboard
                        during training (default is False)
  --page-through PAGE_THROUGH
                        Whether the test meshes should be opened in an interactive
                        session (default is False)
  --result-dir RESULT_DIR
                        The results directory for the tests (default is results)

Based on: Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J. Black. "Generating 3D faces using Convolutional Mesh Autoencoders." European Conference on Computer Vision (ECCV) 2018.