Optical Flow Estimation via Realistic Distraction and Pseudo-Labeling

The patent transforms the input data by applying distraction images and pseudo-labeling techniques, changing the parameters of the training data to include more varied and realistic scenarios. This allows the model to learn from enhanced data without increasing the fundamental complexity of the optical flow estimation system.

The patent creates pseudo-labels by copying and transforming existing ground truth annotations to generate additional training samples. This approach enriches the training data without requiring additional complex system components or increasing system complexity.

The patent implements self-service through pseudo-labeling, where the model generates its own training labels by processing existing data and ground truth annotations. This automated label generation reduces the need for manual annotation while maintaining training quality, allowing the system to serve itself in creating training data.

The patent performs preliminary action by pre-processing ground truth annotations to create pseudo-labels before actual training. This advance preparation of training data reduces the need for extensive manual annotation during the training process, saving time and resources.

The patent changes the parameters of the training data by introducing distraction images and pseudo-labels that simulate real-world variations. This enhances the model's exposure to diverse conditions while maintaining a relatively simple training framework, thus improving robustness without significantly complicating the training process.

The patent introduces distraction images as an intermediary element during training. These distraction images act as mediators that bridge the gap between simple labeled data and complex real-world scenarios, allowing the model to learn robust features without requiring complex training procedures.