This project is a set of graph structure-based scene graph generation. All codes based on PyTorch.

• GCN: Simple Graph Convolution Network. Using 2 GCN layers to update objects and relations separately.
• GraphSAGE: Simple Graph Sample and AggregatE model. Using 2 GraphSAGE layers to update objects and relations separately.
• HGNN: Heterogeneous GNN can handle Object-Object (OO) and Object-Relation-Object (ORO) graphs and uses a custom heterogeneous graph neural network (GNN) for learning and prediction
• TripleGCN: Implemented from Wald et al.'s famous SGPN (CVPR'2020). TripleGCN first generates a <subject, predicate, object> triplet to deal with the features in semantic graphs, and then aggregates the objects & relations features.
• More comming soon...

• Python 3.x
• PyTorch
• NumPy
• Matplotlib

To use this framework, you need to design your own obj_vecs, rel_vecs, and edge_index matrices as following.

obj_vecs: shape (O, D), where O is the number of objects in the scene and D is the dimension of each object feature vector

rel_vecs: shape (T, D), where T is the number of relations in the scene

edge_index: shape (T, 2), edge_index[k] = i, j indicates a triplet < object_vecs[i], rel_vecs[k], obj_vecs[j] >

The code is still being tested. The correctness and validity of the model cannot be guaranteed.

This project is released under the MIT license.