Training a vision model on your own data: a practical playbook
Off-the-shelf models rarely know your products, defects or sites. Here is how we go from raw images to a model you can trust on the line.
By the NIMTECH engineering team7 min read
Start from the decision
A vision project goes well when everyone agrees what the model is for: reject a part, count stock, flag a missing helmet, read a meter. That decision sets the classes, the acceptable error rate and how fast the answer is needed, which in turn decides whether the model runs in the cloud, on an edge device or on a phone.
Collect for variety, not volume
Ten thousand near-identical images teach less than a few hundred that cover different lighting, angles, shifts, cameras and product variants. We start with a small, deliberately varied set, label it carefully and write a labelling guide with examples of every hard case, so labels stay consistent as the team grows.
After the first model is trained, its mistakes tell you what to collect next. This active-learning loop is far cheaper than labelling everything up front.
Evaluate where it will run
Hold back test images from the actual site and conditions the model will face, and never train on them. Report results per class and per condition, not just one overall score, because a model that is excellent on day-shift images and poor at night is not ready.
Ship it with a feedback loop
Deployment is the start, not the end. Log low-confidence predictions, sample them for review, and feed corrected labels back into training on a regular schedule. With monitoring in place you will see drift, such as a new product line or a moved camera, before it shows up as a missed defect.