Depth aware RCNN

Image object detection networks that depend on region proposal networks (RPN) have achieved state-of-art results. As RPN is trained to share convolutional features with the actual classification layers in the network, features learned by the convolutional backbones may have subtle impact on the RPN. A successful approach comes from RGB-D image object detection, where the convolutional layers learn not just RGB features, but also depth features. In this thesis, we study the problem of simultaneously localizing objects as well as estimating their depth. We propose to use one backbone network for two tasks and show that multi-task learning with shared weights can have reciprocating benefits. Our experiments show that when combined with depth prediction in the network, the object detection branch in our model outperforms Faster-RCNN on the challenging KITTI detection benchmark and the Cityscapes dataset. Likewise, the performance of our depth prediction branch is slightly better compared with methods using the same depth prediction architecture.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Zhao, Tianxi

Contributors dc:contributor

• Shi, Honghui

• Object detection
• Depth estimation
• Computer Vision
• RCNN

Statement dc:rights

• Copyright 2019 Tianxi Zhao

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/105082

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/105082