ZOOM Link: https://cmu.zoom.us/j/93785335144

Title: Combining Semantic and Geometric Understanding for Modern Visual Recognition Tasks

Abstract:

For autonomous driving perception, visual data, such as camera image and LiDAR point cloud, consists of two aspects: semantic feature and geometric structure. While usually studied separately, these two properties can be combined and jointly used by a unified framework. In this work, we apply and validate this idea on modern visual recognition tasks. For image panoptic segmentation, we introduce position-sensitive embedding that is able to distinguish instances with similar appearance but at different locations. Such embedding allows a simple single-stage network to generate panoptic segments in a highly efficient way. For LiDAR point cloud detection, we fuse deep semantic feature extracted from pseudo range image with raw geometric information. This additional feature fusion stage significantly improves the detector’s mAP on all categories, outperforming other state-of-the-art approaches.

Committee:

Martial Hebert (advisor)

David Held

Allison Del Giorno