Unsupervised Monocular Depth Estimation via Recursive Stereo Distillation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Unsupervised monocular depth estimation methods often underutilize stereo information during training.
• Existing approaches struggle to fully exploit stereo pairs, limiting monocular depth estimation performance.

Purpose of the Study:

• To propose a novel dual-network architecture for enhanced unsupervised monocular depth estimation.
• To improve the performance of monocular depth estimation by leveraging stereo information more effectively during training.

Main Methods:

• A novel architecture combining a monocular network (Mono-Net) and a stereo network (Stereo-Net).
• Stereo-Net employs a recursive estimation and refinement strategy for accurate depth map prediction.
• A multi-space knowledge distillation scheme transfers expertise from Stereo-Net to Mono-Net.

Main Results:

• The proposed framework achieves superior performance in monocular depth estimation compared to state-of-the-art methods.
• Mono-Net, when trained with Stereo-Net, provides accurate depth estimation with fast runtime during testing.

Conclusions:

• The novel dual-network approach effectively enhances unsupervised monocular depth estimation.
• Knowledge distillation enables a lightweight Mono-Net to achieve high performance by learning from a sophisticated Stereo-Net.