Static Image Training Set for Robust Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer vision systems in real-world manufacturing and warehousing environments often experience random glitches, failing to recognize objects or hallucinating phantom objects, even under consistent viewing conditions.
Innovation Solution
A machine learning technique involving a static image training set with multiple images of the same static object, captured from the same camera position with no changes in the foreground or background, is used to enhance object detection. The images are captured in quick succession or over a longer period with natural changes, and some images may have their properties randomly altered during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple images of the same static object are used in training data, then the robustness of the computer vision system is improved, but the training data size increases
Solution Approach 1:
The patent creates multiple copies of the same static object image to form a training set. Instead of using truly diverse images, the system generates multiple identical or near-identical copies of a single object image, which surprisingly improves system robustness without requiring large amounts of varied training data
Solution Approach 2:
The patent applies random parameter changes to the training images, such as modifying brightness, contrast, color balance, and other image properties. This allows the system to train on multiple variations of the same object without needing multiple actual object images, effectively managing training data size while improving robustness
2Adaptability or versatility
If images with random property changes are applied during training, then the system's adaptability to varying conditions is improved, but the complexity of the training process increases
Solution Approach 1:
The patent applies random property changes to images during the training phase before deployment. By pre-exposing the system to various image variations during training, the system learns to handle diverse conditions without requiring complex real-time adjustments during operation
Solution Approach 2:
The training process dynamically randomizes image properties such as brightness, contrast, and color to create varied training samples from static object images. This dynamic transformation during training helps the system adapt to varying environmental conditions while maintaining a simple static training dataset
Data Source
AI summary
A system includes a camera system configured to capture images of items in a tote. The camera system is configured to interface with an artificial intelligence (AI) system to process images in a variety of ways to support picking and/or placing operations by a robot. Such processing includes planning pick points for items, applying a filter to an image, and planning movement of a robot to perform a picking operation. The camera system further includes an augmented reality tag (ARTag) that is used to calibrate cameras and other devices. The system is configured to train the AI system based on images collected by the camera system wherein the camera system randomly varies the camera settings for each image.