Fig. 3

Feature map visualizations and improved segmentation throughout learning A Examples of feature map visualizations on resnet-101 (for an explanation, see Materials and methods). a An early layer shows activations around the cob shape and the ruler on the right. b The next layer shows more clarified cob shapes with activations mainly at the top and bottom of cobs. c A later layer shows different activations inside the cob. d The latest layer masks the background very well masked from cobs and rulers. B Visualization of the main detection procedure of Mask R-CNN. a The top 50 anchors obtained from the region proposal network (RPN), after non-max suppression. b, c, d show further bounding box refinement and e shows the output of the detection network: mask prediction, bounding box prediction and class label. All images are quadratic with a black padding because images are internally resized to a quadratic scale for more efficient matrix multiplication operations