Large-scale consistensy in visual image elements is essential for figure-ground organization

“

When we perceive a hole, the shape of the hole is convex. However, that area is apparently not perceived as a figure. This is a problem for neuro-computational models to reproduce figure-ground organization based on convexity (and other borderline-based features). I had known this problem for quite a while. When I read the paper about border-ownership sensitive neurons from Rudiger von der Heydt's lab (Zhou et al., 2000, J. Neurosci.), I was intrigued by the fact that they are either contrast-sensitive and -insensitive neurons. I came up with the idea that the long-range interactions between the border-ownership sensitive neurons must reflect the consistency of the surface properties by detecting contrast at individual border segments. I felt it may solve the problem of the hole perception and implemented the idea to our model. We were very excited when we saw, for the first time, the responses of the model to images with complex depth orders (including holes and occlusions). It took us a long while to publish it due to the changes in our professional life. But we are very happy, especially with the development of the "augmented" version of the model. It implements interactions between global and local signal computations. We discussed that these intercations may be mediated by top-down feedback signals.

”