# Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data

**Authors:** Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, Hengshuang Zhao  
**arXiv:** [2401.10891](https://arxiv.org/abs/2401.10891)  
**Date:** 2024-01-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.10891v1/

## Key points

- Depth Anything is a practical solution for robust monocular depth estimation
- The authors scale up the dataset by collecting and annotating large-scale unlabeled data (~62M)
- Two strategies are used to make data scaling-up promising: leveraging data augmentation tools and developing an auxiliary supervision method
- Extensive evaluations demonstrate the impressive generalization ability of the proposed approach
- Fine-tuning with metric depth information from NYUv2 and KITTI leads to new state-of-the-art results
- The authors release their better depth model and depth-conditioned ControlNet on GitHub for further research and applications
- A data engine is designed to automatically generate depth annotations for unlabeled images, enabling data scaling-up to arbitrary scales
- Challenges like occlusions and textureless regions need to be addressed when utilizing large-scale unlabeled data for monocular depth estimation.

## Summary

, , , , 

This work presents Depth Anything, a highly practical solution for robust monocular depth estimation. The goal is to build a simple yet powerful foundation model that can handle any images under any circumstances. To achieve this, the authors scale up the dataset by designing a data engine that collects and automatically annotates large-scale unlabeled data (~62M). This significantly increases the data coverage and helps reduce generalization errors. The authors investigate two strategies to make data scaling-up promising. First, they create a more challenging optimization target by leveraging data augmentation tools. This compels the model to actively seek extra visual knowledge and acquire robust representations. Second, they develop an auxiliary supervision method to enforce the model to inherit rich semantic priors from pre-trained encoders. Extensive evaluations are conducted on six public datasets and randomly captured photos to demonstrate the impressive generalization ability of the proposed approach. Furthermore, fine-tuning with metric depth information from NYUv2 and KITTI leads to new state-of-the-art results. The authors also release their better depth model, as well as a better depth-conditioned ControlNet, on GitHub for further research and applications. In addition, the authors design a data engine that automatically generates depth annotations for unlabeled images. This enables data scaling-up to arbitrary scales by collecting 62M diverse and informative images from eight public large-scale datasets. They use raw unlabeled images without any forms of labels and then annotate them using an initial monocular depth estimation (MDE) model trained on 1.5M labeled images from six public datasets. Despite the advantages of using monocular unlabeled images, it is not trivial to effectively utilize such large-scale unlabeled data due to challenges like occlusions, textureless regions, etc. Overall, this work provides a comprehensive solution for robust monocular depth estimation by leveraging large-scale unlabeled data and incorporating effective strategies for optimization and supervision.