Summary
Smart Info-Media Systems in Asia
2019
Session Number:SS2
Session:
Number:SS2-6
Depth Estimation from a Single Shot Image Using Feature Pyramid Network
Yudai Fukuda, Takuro Oki, Ryusuke Miyamoto,
pp.49-53
Publication Date:2019/9/4
Online ISSN:2188-5079
DOI:10.34385/proc.57.SS2-6
PDF download (5.2MB)
Summary:
Depth estimation is a popular research topic in the field of computer vision. Recent schemes based on deep learning are showing good results for this task, although hand-crafted features and Markov random field were popular several years ago. This paper introduces a feature pyramid network extracting global features from input images into depth estimation, which was originally proposed for object detection. To show the validity of the feature pyramid network, a neural network for depth estimation from a single shot image composed of ResNet-50 and the feature pyramid network was implemented. Experimental results using the KITTI dataset showed that RMSE was improved by about 5% by the proposed scheme with an acceptable decrease of computational speed, resulting in a processing speed of about ten frames per second on a NVIDIA GV100 GPU with 32GB memory.