Learned Stereo Architecture Using Synthetic Data Across Multiple Baselines

The patent implements a universal stereo architecture that can process multiple baselines through a single unified model. The system uses a baseline embedding layer that converts different baseline values into learned representations, allowing the same network weights to adapt to various stereo camera configurations without requiring separate trained models for each baseline.

The system changes the baseline parameter dynamically by feeding it as an input embedding to the network. Instead of training separate architectures for different baselines, the baseline value is transformed into a parameter embedding that modulates the network's processing, enabling a single architecture to handle multiple baseline configurations.

The patent uses synthetic rendered images as copies of real scenes to train the stereo architecture. These synthetic images are generated with known ground truth disparities through computer graphics rendering, providing unlimited training data without requiring time-consuming manual annotation of real images while maintaining sufficient precision for depth estimation.

The system performs preliminary generation of synthetic training data with known ground truth before actual model training. By pre-rendering diverse stereo image pairs with accurate disparity information through graphics rendering, the system eliminates the need for time-consuming real image collection and manual labeling during the training phase.

The patent replaces expensive real image data labeling with synthetic rendered images that provide ground truth disparities automatically through the rendering process. This copying approach using virtual scene rendering eliminates the need for costly manual annotation of real stereo image pairs while maintaining training effectiveness.

The system uses inexpensive synthetic rendered images as disposable training data instead of expensive real labeled images. These rendered images can be generated in unlimited quantities at minimal computational cost, replacing the need for expensive and limited real-world annotated datasets.

The system maintains a single architecture while achieving baseline adaptability through parameter embedding. The baseline parameter is transformed into an embedding vector that is integrated into the network, allowing the same architectural structure to dynamically adapt its behavior based on the specific baseline value without requiring separate models.

The patent introduces a baseline embedding layer as an intermediary between the input images and the main network. This embedding layer translates different baseline configurations into learned representations that guide the subsequent processing, enabling a single architecture to handle multiple baselines through this mediating transformation layer.