3D Keypoint Grouping for Multi-Object Pose Recognition

The patent applies preliminary action by pre-generating multiple learning object models with varied shapes and surface properties, and pre-simulating multiple scenes with different environments before actual recognition tasks. This preparation of diverse training data in advance allows the neural network to be trained on comprehensive variations without requiring extensive data collection during deployment, thus achieving robust recognition while reducing real-time data preparation time.

The patent uses copying by creating multiple virtual copies of object models with different shapes and surface properties through digital manipulation rather than physical object creation. These copied and modified models are then used in simulated scenes to generate training data, eliminating the need to collect vast amounts of real-world data for every possible variation, thereby reducing data preparation time while maintaining recognition robustness.

The patent replaces manual human annotation with automated synthetic label generation by copying and transforming object models programmatically. The system automatically generates correct labels for multiple learning object models and simulation scenes through computational processes, eliminating human annotation errors while maintaining comprehensive training data coverage across various shapes and environments.

The patent substitutes the mechanical process of manual human annotation with an automated computational system. Instead of relying on human operators to attach labels to images, the system uses algorithms to automatically generate ground truth labels from the known parameters of simulated objects, thereby eliminating human error while maintaining versatility in training data generation.

The patent applies universality by designing a recognition model that can handle both single and multiple objects through the same keypoint detection mechanism. The neural network detects keypoints on any number of objects in the scene and the system automatically associates these keypoints with their respective objects, allowing the model to scale from single-object to multi-object recognition without fundamental changes to the implementation, thus achieving multi-functionality.

The patent uses segmentation by dividing the recognition task into independent keypoint detection for each object, followed by clustering and association of detected keypoints into groups corresponding to individual objects. This segmentation approach allows the system to process multiple objects simultaneously by treating each object's keypoints as a separate subset, maintaining implementation simplicity while enabling multi-object capability.