Monocular Terrain Depth Estimation with Classification of Depth Distribution

DOI

Bibliographic Information

Other Title
  • 深度分布の分類を用いた地形画像の単眼深度推定

Search this article

Description

We propose a method for estimating depth from single-view terrain landscapes using convolutional neural networks (CNNs). The existing depth estimation methods using CNNs mainly target at near-distance views (e.g., indoor scenes), and often suffer from large-scale scenes with significant depth range. We attribute this performance degradation to the difficulty of training due to considerable depth variations. In this work, we integrate a CNN for classifying depth distributions inferred from input images on top of depth estimation to alleviate the influence of depth variations during training and inference. We first classify training data into several depth classes and train a depth-estimation CNN for each class. During inference, we obtain a final depth map by calculating a weighted average of outputs from class-wise CNNs, whose weights are obtained as likelihoods inferred by the depth-classification CNN. We attempt more accurate depth estimation for terrain landscapes with this proposed framework.

Journal

Keywords

Details 詳細情報について

  • CRID
    1390577078293136896
  • DOI
    10.11371/wiieej.19.04.0_64
  • ISSN
    27589218
    02853957
  • Text Lang
    ja
  • Data Source
    • JaLC
  • Abstract License Flag
    Disallowed

Report a problem

Back to top