Multiple Instance Feature for Robust Part-based Object Detection

Feature misalignment in object detection refers to the phenomenon that features which re up in some positive detection windows do not re up in other pos- itive detection windows. Most often it is caused by pose variation and local part deformation. Previous work either totally ignores this issue, or naively per- forms a local exhaustive search to better position each feature. We propose a learning framework to mitigate this problem, where a boosting algorithm is performed to seed the position of the object part, and a multiple instance boosting algorithm further pursues an aggre- gated feature for this part, namely multiple instance feature. Unlike most previous boosting based object de- tectors, where each feature value produces a single clas- sication result, the value of the proposed multiple in- stance feature is the Noisy-OR integration of a bag of classication results. Our approach is applied to the task of human detection and is tested on two popular benchm...

• Type: Conference
• Year: 2009
• Where: CVPR
• Authors: Zhe Lin (University of Maryland at College Park), Gang Hua (Microsoft Live Labs Research), Larry Davis (University of Maryland)
• Keywords: Computer Vision, CVPR 2009, Part-based Object Detection