A Single Image Neuro-Geometric Depth Estimation

Abstract

This paper introduces a novel neuro-geometric methodology for single image object depth estimation, abbreviated as NGDE. The proposed methodology can be described as a hybrid methodology since it combines a geometrical and a deep learning component. NGDE leverages the geometric camera model to initially estimate a set of probable depth values between the camera and the object. Then, these probable depth values along with the pixel coordinates that define the boundaries of an object, are propagated to a deep learning component, appropriately trained to output the final object depth estimation. Unlike previous approaches, NGDE does not require any prior information about the scene, such as the horizon line or reference objects. Instead, NGDE uses a virtual 3D point cloud projected to the 2D image plane that is used to estimate probable depth values indicated by 3D-2D point correspondences. Then by leveraging a multilayer perceptron (MLP) that considers both the probable depth values and the 2D bounding box of the object, NGDE is capable of accurately estimating the depth of an object. A major advantage of NGDE over the state-of-the-art deep learning-based methods is that it utilizes a simple MLP instead of computationally complex Convolutional Neural Networks (CNNs). The evaluation of NGDE on KITTI indicates its advantageous performance over relevant state-of-the-art approaches.