Border ownership assignment in real images

Although humans can easily understand from single images the 3D structure of the depicted scene, interpreting a complex scene remains an open research question in Computer Vision. Current computational methods employ both top-down (e.g. semantic labels) and bottom-up (e.g. edges) information. In this work, we address the detection of a visual cue that captures relative depth information, the so-called border ownership, in real images by analyzing intensity patterns near edges and object boundaries. Leveraging on a fast classification technique known as the Structured Random Forest (SRF), we embed two border ownership cues into the SRF: 1) dominant Spectral Patterns and 2) Gestalt-like grouping patterns. Experimental evaluation of the proposed approach over two diverse datasets of real images: a) The Berkeley Segmentation Dataset (BSDS) (200 outdoor images) and b) The NYUDepth (1449 indoor images) shows that our approach is not only more accurate than the previous state-of-the-art, but is able to predict both boundary and border ownership together in real-time: ~0.1s image.