Semi-Supervised Multi-Task Learning with EMA and Pseudo-Labels

The patent applies preliminary action by pre-processing images to generate pseudo-ground truth labels before training. The system creates synthetic training data with predicted depth maps and segmentation masks as pseudo-labels, which are then used for multi-task learning. This preliminary preparation of training data with ground truth equivalents enables the model to learn multiple tasks simultaneously without requiring expensive paired datasets for each task.

The patent uses copying by generating synthetic pseudo-ground truth labels that replicate the structure and format of real ground truth data. The system creates copied versions of training images with synthesized depth and segmentation information, allowing the model to practice multiple tasks on replicated data without needing additional real-world annotated datasets for each task.

The patent applies merging by integrating multiple task-specific modules into a unified multi-task learning framework. Instead of training separate models for each task, the system combines depth estimation, segmentation, and other tasks into a single model that shares common features and learns multiple objectives simultaneously. This consolidation reduces the total number of training parameters while maintaining task-specific performance through multi-objective optimization.

The patent implements universality by designing a multi-functional model that can perform multiple tasks (depth estimation, segmentation, etc.) simultaneously. The unified architecture uses shared feature extraction layers that serve all tasks, reducing redundancy and parameter count while enabling the model to generalize across different tasks. This multi-functional approach improves convergence rate by eliminating the need to train separate specialized modules for each task.