open-mmlab
question about spatial partition strategy and adjacency matrix A
Hi! I have some confusion about the construction of the adjacency matrix A. From the highlighted text and Figure 3 in the paper, it can be seen that spatial partitioning strategy should be based on the coordinates of all joints in a frame to calculate the center of gravity. Then the neighbor set is divided into 3 subsets according to the distance from nodes to the center of gravity and the distance from the root node to the center of gravity. In this case, the subsets of different frames may be different because of different gravity centers. However, I found that A is generated with a fixed center (for example, ntu-rgb+d is always node 20 https://github.com/open-mmlab/mmskeleton/blob/master/mmskeleton/ops/st_gcn/graph.py#L64 ), and the distance calculation is also replaced by hop_dis https://github.com/open-mmlab/mmskeleton/blob/master/mmskeleton/ops/st_gcn/graph.py#L120 (maybe the distance between nodes under the graph structure?). @yjxiong @yysijie Thanks!
Same question haha, but I think that's a fair approximation to use this. However, in these 5 years, is there an issue or PR that answer this?
