Reproduce “Monocular Depth Estimation from a Single Image: Progress and Opportunities” on OpenResearch

About this paper

Muxin Liu, Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Ziyi Yang, Peng Dai, Xiaojuan Qi

Researchers from The University of Hong Kong provide a comprehensive survey of monocular depth estimation, charting its evolution from early learning-based methods to the current era of foundation models. Their work consolidates architectural advancements, training objectives, and data strategies, and includes a unified evaluation of 20 contemporary models across 10 diverse datasets to assess their performance and generalization capabilities.

Key insights

• The emergence of depth foundation models, leveraging large-scale pretraining and diverse data, has significantly advanced zero-shot generalization and robustness in monocular depth estimation.
• Affine-invariant loss functions are crucial for effectively training models on heterogeneous datasets with varying depth annotations, enabling unification of diverse data sources.
• Vision Transformer-based architectures, often initialized with powerful pre-trained encoders, consistently demonstrate superior performance in capturing global context compared to purely CNN-based approaches.

Summary by alphaXiv.