상세정보
학술대회
Monocular Depth Estimation using Improved CSR
- 저자
- 유정재
- 발행일
- 202210
- 출처
- International Conference on Information and Communication Technology Convergence (ICTC) 2022, pp.1912-1916
- 협약과제
- 22IH1700, 인공지능 기반의 사진/회화 융합 및 확장생성 콘텐츠 제작기술 개발, 유정재
- 초록
- Depth-estimation from a single input image can be used in applications such as robotics and autonomous driving. Re-cently, depth-estimation networks using CSR(Channel to Sapce unRolling) [1] has been proposed. The purpose of CSR is to reduce computation in the decoder part of a UNet-typed depth-estimation network. By upsampling at a high magnification, CSR converts thick channel and low resolution data into thin channel and high resolution data. We go one step further and propose a method to increase the accuracy by adding a kind of preprocessing process before the reshape operation during the CSR process. When the proposed method is applied, the amount of computation is slightly increased compared to the case of applying a simple CSR, but the effect of improving the accuracy was confirmed. The proposed method was tested on the NYU-V2 dataset, and it was confirmed that a similar effect was obtained when both a lightweight network and a heavy-weight network were used as encoders.
- KSP 제안 키워드
- Monocular depth estimation, One-step, Single-input, Thin channel, autonomous driving, high resolution data, low resolution