The Graph R-CNN model is divided into four parts:
（1）Target detection, Get proposals that may include targets

（2）RePN(Related Proposal Network): Relational PROPOSAL network.
The role of this network is to sparse the number of objects with complexity O (n), that is, pruning. (If the target detection network detects n proposal, then all possible objects are n*(n-1), that is, the complexity is O (n))

The REPN calculates the correlation of any pair of objects. Pi means the classification probability of the object (object) of the i -i node (object).
φ and are the mapping function (two layers of MLP). The probability of a pair of object classification is maximized, and the two are used as a matrix product to get the SCORE MATRIX. Then do NMS, and the remaining M is that the objects that are meaningful to be connected are right.
The reason why the classification probability is used as the basis of pruning is because some categories are difficult to interact with other categories.

（3）aGCN(Attention Graph Convolutional Network):

This is not clear (very unclear), I wrote it according to my own understanding.
Attention calculation formula (attention of node I):

↓ This is a feature update formula for the relationship (edge):
The current relationship features come from the characteristics of the last characteristics of the previous itself and the two objects it contacted (ie, Subject and Object).

↓ This is the feature update formula of the node (object):
αSKIP represents Skip-Connection, which is available between all Object Node, which is newly added after the REPN. α^skip_ii= 1 (The Attention of itself is 1), the remaining αSKIP should also be calculated according to the above attention.

The formula above is still too abstract. I tried to draw a schematic diagram, and I didn't know if it was right. Maybe I could really understand the code in the future.

↓ Update relationship characteristic schematic diagram

↓ Update object feature schematic diagram (more complicated)
On the left is information from other Object, and the right side is the information from the neighbor.

(4) Category. After the above three steps, each node of Graph has the characteristics of integrating global information. Classification of object categories on the node, classification of the relationship category, and the loss function Crossesntropy.