Learning of Monocular Camera Depth Estimation Using Point-of-view Shots
| Autor: | Byeong-Wook Yoo, Deok-Ho Kim, Taehyuk Kwon, Sunghoon Yim, Won-Woo Lee, Gun-Ill Lee, Jiwon Jeong |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Motion compensation
Image quality business.industry Computer science 3D reconstruction ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION Image stabilization Computational photography Computer Science::Computer Vision and Pattern Recognition RGB color model Computer vision Artificial intelligence business Stereo camera Camera module |
| Zdroj: | 2021 International Conference on Machine Vision and Applications. |
| Popis: | Depth estimation using RGB images plays an important role in many computer vision applications, such as pose tracking for navigation, computational photography, and 3D reconstruction. Depth estimation using a single camera has relatively low accuracy compared to that using a conventional stereo camera system and time-of-flight (ToF) sensor. Nowadays, cameras in smartphone have the optical image stabilization system and can rotate to improve image quality such as image stabilization and image deblur. In this paper, we propose a novel depth estimation technique using a single camera equipped with tilting mechanism. Instead of motion compensation for optical image stabilization, a typical usage of tilting mechanism, we exploit it to capture multi-view images from a single viewpoint, by rotating the camera module in varying angles. The captured images have inside-out views, which is more challenging in depth estimation than out-side in views. We train a network based on structure-from-motion algorithm for depth estimation of a scene to achieve high accuracy depth maps. A synthetic dataset is created from 3D indoor models to train and validate the network. Evaluation results demonstrate that we achieve state-of-the-art performance in depth estimation using a single camera. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|