Pattern-Triggered AR Object Modification With Contour-Anchored Occlusion

The patent applies local quality by making different parts of the virtual object have different occlusion properties. The virtual object uses semi-transparent or translucent materials in certain regions to allow partial visibility of the physical object underneath, rather than complete occlusion. This creates a gradient effect where the virtual object blends more naturally with the physical environment, reducing the uncanny valley effect while maintaining integration.

The patent employs composite materials by combining virtual objects with translucent or transparent material properties with opaque regions. This creates a composite visual effect where the virtual object appears to be partially made of glass or translucent material, allowing the physical object to be seen through certain portions. This composite approach reduces the unnatural appearance of complete occlusion while maintaining the virtual object's presence.

The patent applies segmentation by dividing the computing task into multiple stages. First, a lightweight model performs initial object detection and candidate identification. Then, a more sophisticated convolutional neural network is applied only to the detected candidate objects for detailed analysis and contour determination. This segmented approach reduces the overall computing burden while maintaining high detection accuracy for the final output.

The patent implements preliminary action by performing preprocessing steps before the main convolutional neural network processing. This includes initial object candidate identification, contour detection, and feature extraction that prepare the data in advance. By performing these preliminary actions, the main CNN model receives pre-processed input that requires less computational effort to achieve high accuracy, thereby reducing overall computing resource requirements.