Planar Surface Detection with Unsupervised Neural Networks

The patent inverts the traditional supervised learning approach by using an unsupervised learning scheme that does not require ground truth labels. Instead of training with labeled data and applying to real environments, the method trains and applies simultaneously without labels, resolving the contradiction between accuracy and adaptability.

The method enables the system to self-train and self-evaluate without external ground truth labels. The unsupervised learning scheme allows the neural network to automatically learn plane parameters from raw depth images and evaluate its own performance through consistency checks, making the system self-sufficient in real-world applications.

The system performs self-training without requiring external labeled data. The unsupervised learning framework enables the neural network to automatically learn from unlabeled depth images, eliminating the need for complex ground truth label acquisition, management, and annotation processes.

The patent introduces an intermediate self-evaluation mechanism that acts as a mediator between the neural network and the environment. This self-evaluation process using consistency checks replaces the need for ground truth labels, simplifying the data preparation process while maintaining detection capability.