Monocular depth estimation for spacecraft: combining relative and scale information

Abstract

This paper proposes a novel object-specific depth estimation method for space applications, combining multi-scale feature fusion, transformer improvements, and frequency-based refinement in the Discrete Cosine domain (DCT). It addresses challenges like high contrast, noise, and limited compute by leveraging segmentation maps to focus computation. The method avoids iterative full-image processing and aims for real-time, robust performance. Benchmarks are performed on spacecraft such as LRO and Dream Chaser, showcasing state-of-the-art depth accuracy, geometric learning and structural reconstruction showcasing sub 0.5 meter accuracy.