Currently, most point cloud based detection pipelines are focused on producing high accuracy results while requiring significant computational resources and a high-end GPU. Our research explores how to reduce the computational overhead by improving a key element of detection: bounding box regression. We demonstrate a fast and iterative method of bounding box regression which has significant runtime performance advantages over existing leading methods. The iterative structure of our method also gives the system control over the tradeoffs between accuracy and speed at runtime. We furthermore integrate our bounding box regression method into an existing detection pipeline and motivate additional research into how the first stage of the pipeline can be modified to take better advantage of the performance characteristics of our bounding box regression method.

Committee:

Simon Lucey (advisor),

Katerina Fragkiadaki,

Allie Chang