Synthetic dataset generation for object-to-model deep learning in industrial applications

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Large image datasets are vital for deep learning (DL) success in classification and detection.
• Industrial use-cases, like warehouse product identification, suffer from scarce data, hindering DL deployment.

Purpose of the Study:

• To develop a computer vision system for detecting and identifying supermarket products in a warehouse.
• To demonstrate a framework for data synthesis to create an end-to-end deep learning pipeline for industrial applications.

Main Methods:

• Generated a synthetic dataset using 3D models created via photogrammetry of real-world objects.
• Trained an InceptionV3 convolutional neural network on 100,000 synthetic images across 10 classes.
• Utilized automatic pixel annotation during image generation to eliminate manual annotation costs.
• Trained a RetinaNet detector on the synthetic, annotated images.

Main Results:

• Achieved 96% accuracy on a real-world test set using the trained InceptionV3 model.
• The developed RetinaNet detector accurately localizes and classifies products in real-time.
• Data synthesis significantly reduced the bottleneck of manual data creation and annotation.

Conclusions:

• Data synthesis from 3D models is an effective solution for creating industrial datasets for deep learning.
• The proposed framework enables the development of efficient computer vision systems for warehouse automation.
• This approach accelerates the deployment of deep learning in specialized industrial settings.