This paper focuses on the advancements and methods developed during the RoboDepth Challenge, an academic competition aimed at facilitating and promoting robust out-of-distribution (OoD) depth estimation. This challenge was crucial for safety-critical applications, given the need for accurate depth estimation under diverse conditions, including adverse weather, sensor failure, and noise contamination. Participants proposed solutions involving spatial and frequency-domain augmentations, masked image modeling, image restoration and super-resolution, adversarial training, diffusion-based noise suppression, vision-language pre-training, learned model ensembling, and hierarchical feature enhancement. The goal was to improve depth estimation’s robustness, enabling more reliable predictions under diverse real-world conditions.

 

Publication date: 27 Jul 2023
Project Page: https://robodepth.github.io
Paper: https://arxiv.org/pdf/2307.15061.pdf