Instance Segmentation, Body Part Parsing, and Pose Estimation of Human Figures in Pictorial Maps

In recent years, convolutional neural networks (CNNs) have been applied successfully to recognise persons, their body parts and pose keypoints in photos videos. The transfer of these techniques artificially created images is rather unexplored, though challenging since are drawn different styles, proportions, levels abstraction. this work, we study problems on the basis pictorial maps where identify included human figures with two consecutive CNNs: We first segment individual Mask R-CNN, then parse estimate poses simultaneously four UNet++ versions. train CNNs a mixture real persons synthetic compare results manually annotated test datasets consisting figures. By varying training CNN configurations, were able improve original R-CNN model achieved moderately satisfying extracted may be used for animation storytelling relevant analysis historic contemporary maps.