RainforestDepth: Monocular Depth Estimation Targeting Rainforest Environments

Abstract

The primary objective of this paper is to introduce a new monocular depth estimation (MDE) model targeting under-represented environments using a novel dataset combining synthetic and real images. The proposed model is small and fast to allow use for UAS navigation and data collection in rainforest environments. Prior works on MDEs target outdoor environments while focusing on urban, ground-level viewpoints due to interest in self-driving or autonomous package delivery applications and data availability. However, under-represented environments, such as rainforests, can benefit from targeted, environment-specific MDEs because existing general MDEs cannot adapt to extreme environmental differences, leading to high error rates. Our model is trained using a distinct rainforest dataset that combines images generated using a synthetic dataset pipeline and depth images collected from aerial robot deployments in the Children’s Eternal Rainforest in Costa Rica. The proposed model will allow for improved rainforest navigation without using expensive LIDAR sensors and can improve the navigation of a UAS in rainforest environments by providing more accurate and useful measurements for object manipulation, such as leaf sampling. Our model outperforms MiDaS across the board and has over a 75% improvement, specifically in the relative error metrics, while maintaining a low runtime. The resulting model matches the performance of state-of-the-art monocular depth estimation models designed for common environments, i.e., urban and indoor environments, and outperforms them when used in a rainforest environment.