Training image recognition models

Who it is for

Companies that need to recognise in images what off-the-shelf tools cannot.

What it solves

Off-the-shelf tools cannot recognise what matters in specific data.

Our role

Data preparation, model training and evaluation, and export

What was built

• Preparing and checking training data
• Training and evaluating models
• A register of experiments
• Exporting models for a server or a phone

Technologies

Python, PyTorch, ONNX

When it makes sense

Off-the-shelf tools handle common tasks, but not always one company’s data. A custom model can then be trained. It can learn to recognise many kinds of things, such as people in camera footage, product defects or document types.

How we work

• Data. The dataset is built from the same sources each time, so a result can be repeated.
• Training. We start from a widely used network architecture and extend the data with image augmentation.
• Evaluation. Results, thresholds and the course of each experiment go into a register, so models can be compared.
• Export. A finished model comes as a package with its metrics, thresholds and checksums, in ONNX for a server or LiteRT for a phone.