The traces of Figure 5 were generated using a very local 5-pixel neighborhood in searching ahead. This works better in a relatively noise-free domain, such as the high-resolution photography. However, when noise can interrupt a contour, the network must have a larger perspective to continue past the noise. Some other filters we plan to try are line-extension fields, analogous to those recently identified in complex cells.
Additionally, there are many basic engineering decisions in selecting and/or appropriately weighting training data, when many neighborhoods along the contour are redundant, and a few key cases capture the essence of the contour in its overall environment.
This tracing model shows promise. On straightforward contours, with 20 initial pixels of training data, contours can be followed continuously for hundreds of pixels. What remains to be measured is how well this squares with a ground-truth of an expert's delineation, and over how complex a landscape a network can be adequately trained.
A further extension of this work is into contour identification across the 3D volume composed of many parallel slices. We plan to explore contour extension on adjacent layers, without further training. And when slices are sufficiently well registered, several traced layers could analogously be used to propagate the contour identification to succeeding layers.