Data Handling

We store the data for each training sample or a validation sequence in HeteroData object wrapped in a Batch object from pytorch geometric library.

Nodes of the graph

Each sample has two sets of nodes: "cloth" for garment nodes and "obstacle" for body nodes.

Each of this node types is a dictionary inside the Batch object that stores all the data for this node type (see picture below). This data can be accessed with sample[NODE_TYPE] where NODE_TYPE is either "cloth" or "obstacle".

Edges of the graph

The Batch object also stores data for the edges of the graph which can be accessed with sample[SOURCE_NODES, EDGE_TYPE, TARGET_NODES].

The batch built by the dataset only has a set of mesh edges (extracted from the garment mesh) accessed by sample["cloth", "mesh_edge", "cloth"] and several sets of coarse edges (sample["cloth", "coarse_edgeX", "cloth"]).

During the forward pass of the Model we also add proximity-based world edges (named body edges in the paper): both direct (sample["cloth", "world_edge", "obstacle"]) and inverse (sample["obstacle", "world_edge", "cloth"]).

Dataset's wholeseq parameter

wholeseq parameter of the Dataset, controls whether the dataset returns a whole sequence of body poses (if True) or only a small window of the sequence (if False).

• wholeseq=False is used for training
• wholeseq=True is used for validation and testing

Batch size

As the HOOD model is locally applied to each node of the input graph, it can be trained with batch_size=1 (one pose sequence in a batch).

Because of that, we only use batch_size=1 in this repository. Larger batch sizes are not supported.