The ground truth can be biased by humans inference. That is, humans are able to detect real world objects and ignore existing contourns that are not relevant. I do not know how the classifier learns that by using low level features. I guess it is confused sometimes and this may degrade its performance. I am not sure about abstract art either. Maybe using synthetic images coming from a 3D modeller in which both, humans and computers, know where the contourns are could be a solution to this.

The goal of creating a pure low level boundary detector that should label regions in the same way as a human would is ambitious. Humans probably make use of some top-down knowledge when marking boundaries on pictures, so I am not sure if a gradient based approach like this one will ever achieve human-like results. It would be interesting to see a comparison of their algorithm against human-labeled images of e.g. totally abstract art (which the human has never been exposed to and which does not carry any recognizable meaning in it). I believe that such a comparison will have a significantly smaller gap between human and machine since the human can't take as much advantage of his inherent knowledge of what constitutes an object.

please post you thoughts on "Learning to Detect Natural Image Boundaries Using Local Brightness, Color, and Texture Cues"