GPI Seminar Series: Jordi Pont-Tuset
This contribution brings the potential of deep Convolutional Neural Network (CNN) architectures to the field of perceptual grouping. We present Convolutional Oriented Boundaries (COB), a technique that produces multiscale oriented contours and region hierarchies starting from generic image classification CNNs.
COB is computationally efficient, because it requires a single CNN forward pass for contour detection and uses a novel sparse boundary representation for hierarchical segmentation; it gives a significant leap in performance over the state-of-the-art, and it generalizes very well to unseen categories and datasets. Particularly, we show that learning to estimate not only contour strength but also contour orientation provides more accurate results. We also conclude that our architectures do not require contour globalization, which was one of the speed bottlenecks in existing approaches. We perform an extensive experimental validation on BSDS, PASCAL Context, PASCAL Segmentation, and MS-COCO, showing that COB provides state-of-the-art oriented contours, region hierarchies, and object proposals.
Jordi Pont-Tuset is a post-doctoral researcher in Prof. Luc Van Gool’s ETHZ vision group. Previously, He did an 8-month internship at Disney Research, Zürich, under the supervision of Prof. Aljoscha Smolic, visited Prof. Jitendra Malik’s vision group in UC Berkeley, and collaborated with the startup Fezoo. I am a mathematician, engineer, and PhD in computer vision by UPC Barcelonatech under the supervision of Prof. Ferran Marques.