Closed YanYan0716 closed 6 years ago
In our implementation, we followed the idea of separable convolutions. We used a common 1d convolution to process the temporal information (a joint track) and a graph convolution to process spatial information (a human skeleton in a frame).
You can also process them simultaneously. Please refer to https://github.com/yysijie/st-gcn/blob/master/net/utils/tgcn.py. ConvTemporalGraphical
is the graph convolution layer we used. You can assign t_kernel_size
parameter to modify the kernel size in the time dimension. The default value is 1.
thanks a lot, i will see the code again.
@yysijie @yjxiong hello, thank you for your sharing about the network, as a new student in this field, I want to ask you "how to add the temporal information in the graph ? " in the graph.py, i only saw the adjacency matrix about the 18 points in openpose, but i don't know your method to add the temporal information in the graph, could you give me some advice, of course as a beginner,i am looking for your reply, thanks again