Machine Learning Camera Calibration Saddle Point Detection

The training process is segmented into three distinct phases: Phase 1 trains the network to detect presence/absence of pre-determined points using binary labels; Phase 2 refines position estimation using coordinate labels; Phase 3 applies adversarial training with modified images to improve robustness. This segmentation allows the complex training to be broken down into manageable stages, each building on the previous phase while progressively improving detection accuracy.

Adversarial training is performed as a preliminary action in Phase 3, where images are pre-modified with various distortions, rotations, and lighting changes before training. This preliminary exposure to diverse conditions prepares the network to handle real-world variations more effectively, improving reliability while managing training time by structuring the adversarial training as a dedicated phase rather than continuous refinement.

The three-phase training approach maintains continuity of useful action by ensuring each phase builds directly on the previous phase's results. Phase 1 establishes basic detection capability, Phase 2 refines position accuracy, and Phase 3 enhances robustness - each phase's output becomes the input for the next, creating a continuous improvement trajectory that optimizes calibration quality while managing execution time through progressive refinement rather than repeated full training cycles.