Hi Kotaro -- many thanks for reading my post and I am excited to see what you did with continuous plots. I also plan to use the output tensor (whose elements consist of 1 - background, 2 - axes, 3 - axes labels, 4 - data points) of the modified ResNet-22 net to identify different plot components that are associated with certain pixel positions. For example, a singe data point can be represented by the average of its occupied pixel positions. Similarly, a second neural net can be implemented to recognize the average pixel positions of any axes label. I also look forward to discussing with you further in the future.