Open 7oud opened 1 year ago
Hi 7oud, you are right, the Fig is somewhat misleading. So in general the connections (orginally as a spiral on the mesh, in our case as a loop around the contour) are ordered, hence the arrows in the Fig, but they still have a connection to all vertices in the same time frame plus one connection to the neighboring frame. We decided to have it fully connected but this can be changed.
Note that in the MICCAI implementation there is also a small bug in terms of how the sequence is ordered. If you change the following line in the nets/CNN_GCN.py:
def create_graph(self, num_nodes: int) -> np.ndarray:
adjacency = []
for ii in range(num_nodes):
x = list(range(num_nodes))
#x.insert(0, x.pop(ii)) changed with the line below
x=np.roll(x,-ii)
adjacency.append(x)
adjacency = np.array(adjacency)
Then the order is acutally reflecting the loop. We did some experiments and realized that the final accuracy is not significantly influenced but definetely the performance on how fast the model converges.
From the figures in paper, the vertex
v1
is connect tov2
andv0
in the ED, and there is another edge betweenv1
and its according vertexv1'
in theES
. But after I read the code, it seems thatv1
is connected with all the vertices in the same image,ED
andv1'
inES
. I understand right?