Learning Geocentric Object Pose in Oblique Monocular Images

TL;DR

This paper introduces a deep learning method to estimate geocentric object pose from oblique monocular images, improving localization and alignment in Earth observation tasks without relying on monocular depth.

Contribution

The paper presents a novel encoding of geocentric pose and a deep network trained with lidar supervision to estimate object height and orientation from oblique images.

Findings

Enhanced accuracy in object localization and image alignment.

Effective extension of semantic segmentation datasets for satellite imagery.

Public availability of data and code for reproducibility.

Abstract

An object's geocentric pose, defined as the height above ground and orientation with respect to gravity, is a powerful representation of real-world structure for object detection, segmentation, and localization tasks using RGBD images. For close-range vision tasks, height and orientation have been derived directly from stereo-computed depth and more recently from monocular depth predicted by deep networks. For long-range vision tasks such as Earth observation, depth cannot be reliably estimated with monocular images. Inspired by recent work in monocular height above ground prediction and optical flow prediction from static images, we develop an encoding of geocentric pose to address this challenge and train a deep network to compute the representation densely, supervised by publicly available airborne lidar. We exploit these attributes to rectify oblique images and remove observed object parallax to dramatically improve the accuracy of localization and to enable accurate alignment of multiple images taken from very different oblique viewpoints. We demonstrate the value of our approach by extending two large-scale public datasets for semantic segmentation in oblique satellite images. All of our data and code are publicly available.