Binocular Depth Estimation via Parallax and Confidence Segmentation

The patent segments the prediction output into two distinct components: a parallax image providing depth information and a confidence score image providing reliability information. This segmentation allows the system to process and evaluate different aspects of prediction quality separately, enabling improved accuracy in critical areas without sacrificing overall prediction speed.

The confidence score image serves as a feedback mechanism that indicates the reliability of each pixel's depth prediction. By providing this feedback information, the system can identify areas with low confidence (such as small obstacles) and apply corrective measures or further processing specifically to those regions, thereby improving prediction accuracy where it matters most while maintaining efficient processing elsewhere.

The system performs self-evaluation by generating confidence score images that automatically assess the reliability of its own predictions. This self-service mechanism allows the deep learning network to identify potential errors without requiring external validation or complex verification systems, thereby improving reliability while adding minimal complexity.

The confidence score image acts as an intermediary between the raw prediction and the final decision-making process. It provides a layer of evaluation that bridges the gap between the simple deep learning output and the need for reliable, error-free predictions, enabling the system to handle uncertainty without requiring fundamentally more complex architecture.