3D Point Cloud Augmentation via Randomized Noise and Transformations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Creating and augmenting a 3D objects training dataset based on a small number of models is challenging due to the cost-inefficiency and complexity of conventional methods, requiring specialized hardware and software.
Innovation Solution
A method that involves accessing a 3D point cloud representation, applying an augmentation routine with random selection of noise addition, geometric transformation, and degradation, and adding the augmented point cloud to the training dataset, using techniques like White Gaussian noise, geometric transformations, and degradation to generate diverse training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional methods are used to augment 3D objects training dataset, then the dataset can be expanded, but the cost and complexity increase significantly
Solution Approach 1:
The patent applies digital copying by creating virtual copies of 3D objects through point cloud representations. Instead of physically manufacturing multiple objects, the system digitally replicates objects by applying transformations (rotation, translation, scaling) and adding noise to existing 3D models, thereby expanding the training dataset without requiring additional physical hardware or complex manufacturing processes.
Solution Approach 2:
The patent changes parameters of the 3D point cloud data by applying random transformations such as rotation angles, translation vectors, and scaling factors. It also modifies parameters by adding noise to the point cloud coordinates, which creates variations in the training data while using the same underlying 3D model, thus expanding the dataset efficiently.
2Quantity of substance
If conventional methods are used to augment 3D objects training dataset, then the dataset can be expanded, but the cost increases
Solution Approach 1:
The system creates digital copies of 3D objects through point cloud processing, eliminating the need for expensive physical manufacturing, scanning, or printing of multiple objects. The copying process uses software-based transformations and noise addition, which are computationally inexpensive compared to conventional physical augmentation methods.
Solution Approach 2:
The patent uses computationally inexpensive, disposable operations such as adding random noise and applying simple geometric transformations to generate training data. These operations are cheap and can be performed repeatedly without requiring expensive equipment, making the dataset augmentation process cost-effective.
3Quantity of substance
If conventional methods are used to augment 3D objects training dataset, then the dataset can be expanded, but the process becomes cumbersome
Solution Approach 1:
The system performs simple digital copying operations on point cloud data, such as creating copies and applying transformations. This automated digital copying process is much simpler and more efficient than conventional manual methods, requiring only software execution rather than complex physical manipulation or specialized equipment operation.
Solution Approach 2:
The augmentation process is automated and self-service, where the system automatically generates training data by applying transformations and noise to input 3D models without requiring manual intervention. The process self-manages the entire augmentation workflow, from input processing to output generation, making it operationally simple and efficient.
Data Source
AI summary
A 3D objects training dataset is augmented by accessing a 3D point cloud representation of an object and by applying an augmentation routine on the point cloud to generate an augmented point cloud. The augmentation routine comprises randomly selecting an execution order of at least one of (i) adding a noise to the point cloud, (ii) applying a geometric transformation on the point cloud and (iii) applying a degradation on the point cloud. The randomly selected execution order of these operations on the point cloud is applied, and the augmented point cloud is added to the objects training dataset. A machine learning algorithm (MLA) is trained by inputting the 3D point cloud representation to generate an output, comparing the output with an expected label associated with the point cloud representation to determine a measure of error on the output, and iteratively adjusting various weights associated with nodes of the MLA.