Vision & AI: Faster Dataset Generation
Training artificial intelligence: too time consuming?
In the manufacturing industry, vision technologies can play a important role. From detecting, counting or classifying items to detecting defects, vision combined with deep learning provides fast, robust and accurate results.
But the effectiveness of this technology depends on the quality and size of the datasets on which it was trained. And traditionally, generating these datasets is a very time-consuming process. It requires extensive manual annotations, which can delay the implementation of new systems.
We investigated new workflows to speed up the process of generating image datasets. Our research resulted in a strong 'proof of concept' that rapid development of an AI model is feasible. This is an important step in making the deployment of innovative vision technologies more accessible to the industry.
Conveyer Dataset Generator
We investigated using a conveyer for automatic dataset generation. We developed a sensor tunnel mounted above a conveyor system and equipped with a variety of sensors, including advanced 3D sensors. Objects placed on our 'Conveyor Dataset Generator' were automatically recorded and annotated while passing underneath the sensors.
We tested our conveyer on a wide variety of objects and purposes, such as an oriented bounding box detector and detecting defects in profiles with an anomaly detection approach. The process of using the conveyer is fast and straightforward. During our tests, we were able to generate a dataset of about 100 images in just 30 minutes.
Synthetic Data for Object Counting
Additionally we investigated the potential of generating synthetic images to train an AI model. Starting with only a CAD File we used computer graphics software to automatically create realistic looking scenes and produce photorealistic synthetic data to train an AI model. We added only 10 real images with annotations to the final dataset for training the AI algorithm.
The result? It takes less than a day (6 to 8 hours on a Laptop) to generate 10000 images. Based on these datasets, we trained robust AI models for object counting, for example counting stappled profiles on a pallet.
New workflows: an AI-model in 1 day
The aim of our research was demonstrating that the labor-intensive process of manually annotating data can be mitigated. In our 'proof of concept' it only takes one day to train a full AI model.