Fast Data Generation for Training Deep-Learning 3D Reconstruction Approaches for Camera Arrays.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Neural networks are increasingly used for depth reconstruction from multi-camera arrays.
• Existing training datasets often simulate fixed camera configurations, limiting adaptability.
• Supervised learning for depth reconstruction requires extensive ground truth data.

Purpose of the Study:

• To develop a versatile virtual data generator for creating large-scale training datasets for depth reconstruction.
• To enable training on diverse multi-camera array configurations, including wide-baseline setups.
• To address the limitations of current datasets in simulating realistic and varied scenarios.

Main Methods:

• A purely virtual data generator was developed, capable of adapting to various camera array sizes and inter-camera distances.
• The generator creates virtual scenes with random objects, textures, and user-defined parameters (e.g., disparity range, image resolution).
• Generated data, though unrealistic, incorporate challenges like thin elements and random texture assignment to improve model robustness.

Main Results:

• The virtual data generator successfully created large, adaptable training datasets.
• Experiments focused on wide-baseline configurations, demonstrating the generator's utility for challenging scenarios.
• Validation using established deep-learning and depth reconstruction algorithms confirmed the effectiveness of the generated datasets.

Conclusions:

• The proposed virtual data generator is a valuable tool for creating diverse training data for depth reconstruction algorithms.
• The approach effectively supports the training of models for various camera array configurations, particularly wide-baseline systems.
• The generated datasets enhance the robustness and performance of depth reconstruction models.