Scaling for Multimodal 3 D Object Detection

We investigate two methods for scalable 3D object detection. We base our approach on a recently proposed template matching algorithm [8] for detecting 3D objects. First, we demonstrate that it is possible to gain significant increase in runtime performance of the algorithm at almost no cost in accuracy by quickly rejecting most regions of the image with low-resolution templates. Second, we investigate an implicit part-based model that uses fixed-sized template dictionary and a Generalized Hough Transform framework to detect objects. We present results on two separate datasets that we collected using the Kinect sensor.