AR Toolkit for Cleaning Robot Image Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manually labeling objects in images for training deep learning systems, such as those for cleaning robots, is labor-intensive and time-consuming, requiring a large number of training samples.
Innovation Solution
A computer-implemented method that uses an augmented reality toolkit to overlay a virtual object on a real-world object, allowing for automatic identification and labeling of the object across multiple views and orientations, enabling the generation of a trained recognition module to recognize the object in additional images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to label objects in images for training deep learning systems, then labeling accuracy can be ensured, but the process becomes labor-intensive and time-consuming
Solution Approach 1:
The patent uses a single manually labeled image as a template or copy reference. The labeling system automatically applies this template to generate labels for multiple similar images, copying the labeling pattern rather than manually labeling each image individually. This maintains consistency and accuracy while dramatically reducing time consumption.
Solution Approach 2:
The patent performs preliminary manual labeling on just one representative image to establish the labeling template. This preliminary action on a single image enables automatic labeling of numerous other images, converting what would be repetitive manual work into a one-time setup followed by automated processing.
2Reliability
If a large number of training samples are collected for deep learning systems, then model performance improves, but the labeling process becomes increasingly labor-intensive
Solution Approach 1:
The system copies the labeling template from a single manually labeled image and automatically applies it to generate labels for large numbers of training images. This enables efficient generation of extensive training datasets without proportionally increasing manual labeling effort, thus improving model performance while maintaining labeling productivity.
Solution Approach 2:
The labeling template created from one image serves as a universal pattern that can be applied to multiple different images. This multi-functional approach allows a single manual labeling effort to produce labels for numerous training samples, enabling large-scale dataset creation with minimal additional manual work.
3Measurement precision
If multiple views and orientations of objects are captured for comprehensive training, then recognition accuracy across various conditions improves, but the number of images requiring manual labeling increases
Solution Approach 1:
The system captures multiple images of objects from various views and orientations, then uses the labeling template copying approach to automatically label all these diverse images. This enables comprehensive training data collection covering multiple scenarios without manually labeling each image, maintaining high recognition accuracy while reducing the burden of labeling large quantities of images.
Data Source
AI summary
A computer-implemented method for labeling images includes capturing, using an augmented-reality enabled device, a first set of images that include views of a first object; at the augmented-reality enabled device, for each of the first set of images, identifying the first object and generating a bounding box that is associated with the first object; receiving an input providing a first label for the first object; and at the augmented-reality enabled device, for each of at least some of the first set of images, associating the first label with the first object bound by the bounding box.