Camera Posture Estimation Using Vanishing Points on a Unit Sphere

The patent replaces conventional arc detection algorithms with a machine learning model (neural network) to estimate vanishing points directly from images. This substitution enables accurate vanishing point estimation even in challenging environments with blurred building contours, while the model can be deployed on standard computing hardware without requiring specialized mechanical or optical devices.

The patent performs preliminary actions by pre-training the machine learning model offline with large datasets. Once trained, the model contains learned features and patterns that enable accurate vanishing point estimation during runtime without requiring heavy computational resources for training. This separates the computationally intensive training phase from the deployment phase, allowing accurate inference on devices with limited power.

The patent changes the input parameters and features considered by the estimation system. Instead of relying on arc detection parameters that require clear building contours, the machine learning model learns from diverse image features including texture, color, and structural patterns that remain reliable even when building contours are blurred. This enables the system to adapt to various environmental conditions while maintaining reliable vanishing point estimation.