Speaker: 

Fuxin Li (CSE Postdoc), School of Computational Science and Engineering

Title:

Semantic Image Segmentation by Ranking Multiple Figure-Ground Segment Hypotheses

Abstract:

The goal of semantic segmentation is to recognize objects in images and classify each pixel into a particular category or background. It is significantly more challenging than more conventional image classification and object detection problems. Different from most other approaches that mainly rely on local cues from pixels or superpixels, our approach starts from multiple binary segmentations that capture important global cues, such as the shape of an object.

These binary segmentations, generated by the unsupervised Constrained Parametric Min-Cut (CPMC) algorithm, cover a spectrum of different locations and segment sizes. In the learning phase, the segmentations are first filtered with a global "objectness" filter, then fed into a kernel-based learning framework that continuously predicts the overlap of each segmentation with each particular object category. Finally, cues from multiple highly-ranked segmentations are used to determine the classification of each pixel. From 2009 to 2011, this approach has won the prestigious PASCAL VOC Segmentation Challenge three times in a row.