Learning Object Detectors on the Continuous Viewsphere: Kun He, BU

Starts:
2:00 pm on Friday, February 7, 2014
Ends:
3:00 pm on Friday, February 7, 2014
Location:
MCS 148
Abstract: In this work, we focus on the problem of joint object detection and pose estimation in images. Our goal is to learn a discriminative object detection model that is smoothly and continuously parameterized by the viewpoint. This is in contrast to most existing multi-view object detection approaches that discretize the viewsphere and treat the problem of viewpoint estimation as a multiclass classification problem. In order to smoothly parameterize our model using viewpoint, we adopt the concept of multiplicative kernels originally proposed by Yuan et al. For efficient multi-view inference, we propose several modifications to the Efficient Subwindow Search framework of Lampert et al. Preliminary results on standard datasets indicate that both the localization and viewpoint estimation performances of our method are on par with state-of-the-art approaches. In addition, I will describe ongoing extensions of this work, including feature selection in the multiview setting and cascade inference.