Object Detection using Unsupervised Parts-based Attributes

This paper presents a new approach to parts-based object detection. Objects are described using a spatial model based on its constituent parts. Unlike most existing methods, the parts are discovered in an unsupervised manner from training images with only object bounding boxes provided. The association between parts is modeled using boosted decision trees that allows arbitrary object-part configurations to be maintained. Experimental results on the very challenging Pascal Visual Object Classes Challenge dataset validate our approach.


"Object Detection using Unsupervised Parts-based Attributes"
Santosh K. Divvala, Larry Zitnick, Ashish Kapoor, Simon Baker
Technical Report TR-11-10, CMU, 2010
(PASCAL VOC 2010 Workshop)
[Paper] [Presentation]