Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Monocular depth estimation requires large, diverse datasets.
• Existing datasets have inherent biases and acquisition challenges.
• Combining disparate datasets is difficult due to annotation incompatibilities.

Purpose of the Study:

• To develop tools for effectively mixing diverse datasets for monocular depth estimation training.
• To establish a robust training methodology invariant to scale and range variations.
• To improve the generalization capabilities of monocular depth estimation models.

Main Methods:

• Developed a robust training objective invariant to depth scale and range.
• Utilized principled multi-objective learning for combining heterogeneous data sources.
• Incorporated pretraining of encoders on auxiliary tasks.
• Experimented with five diverse datasets, including 3D films, and employed zero-shot cross-dataset transfer for evaluation.

Main Results:

• Mixing complementary datasets significantly enhances monocular depth estimation performance.
• The proposed approach demonstrates superior generalization across unseen datasets.
• Achieved state-of-the-art results in monocular depth estimation.

Conclusions:

• Combining diverse datasets with proposed methods is crucial for advancing monocular depth estimation.
• The developed tools and training strategies enable effective utilization of varied data sources.
• This work sets a new benchmark for monocular depth estimation accuracy and generalization.