Computer Vision for Business: Practical Uses and Honest Limits

Computer vision is the branch of AI that lets a computer "see" what's in a photo or video: recognise objects, count them, and mark where they are. It's increasingly within reach, since many businesses already have CCTV cameras or decent phones. But there's a real gap between a smooth demo and a system you can rely on every day. This article covers realistic uses, the data you'll need, and the limits to understand before you start.

What computer vision can do

A few task types cover most needs. Classification answers "which category does this image belong to", for example a good product versus a defective one. Object detection marks each item with a box and a label, for example three cars and one truck in a single photo. From detection, a system can count items or track how they move over time.

Keep in mind that a model only recognises the categories it has been taught. A model trained on cars and trucks won't automatically recognise motorbikes, and a model trained on one type of packaging may not recognise it after a redesign. Every new category means new data and retraining.

Realistic use cases

• Counting vehicles entering a car park, drive-thru, or warehouse gate, to see when it gets busy.
• Checking store shelves, for example spotting empty sections or products placed in the wrong spot.
• Inspecting products on a production or packing line, flagging items that are cracked, dented, or labelled crookedly.
• Counting goods in a photo, such as sacks or boxes in a stack, as a cross-check against stock records.

Notice what these have in common: the objects are distinct, the camera angle can be controlled, and an occasional mistake isn't fatal because a person still checks the results. Uses like these make the best first project.

The data you'll need

Computer vision models are usually retrained from an existing model, a process called fine-tuning. For that you need photos from your business's real conditions, not sample images from the internet.

1. Collect photos or video clips from the camera you'll actually use, with the same angle and distance.
2. Make sure there's enough variety: morning, midday, night, different weather, busy and quiet periods.
3. Label every object, meaning a box around each item plus its category. This takes time and has to be consistent.
4. Give every category plenty of examples. Categories that rarely appear will be hard for the model to recognise.
5. Set aside some images for validation, testing the model on pictures it never saw during training.

A real example: our vehicle detection prototype

As a practice project, we fine-tuned a YOLO11n model to recognise three kinds of vehicle in traffic photos: buses, cars, and trucks. It was trained for 50 epochs on a CPU, then used to predict new images at a 0.5 confidence threshold.

Here are the results, as they are. The overall mAP50 on the validation data was 0.603. Put simply, mAP50 measures how accurately a model finds and boxes objects, on a scale from 0 to 1. But that figure comes from only 20 validation images, far too few to treat as a guarantee. The bus class was also very weak, with an mAP50 of just 0.098, because the validation data contained only 6 buses. The lesson is clear: categories with few examples will be recognised poorly, however good the model is.

That's why we treat this project as a prototype and a learning exercise, not a system ready for business use. A real deployment would need far more, and better balanced, data from the actual cameras.

Limits to understand

• When conditions change, accuracy drops. A model that does well at noon can struggle at night, in rain, or after the camera shifts.
• Hidden objects are hard. Vehicles or goods that overlap are often missed or counted twice.
• Accuracy figures depend on the test data. Always ask how many images it was tested on, and under what conditions.
• Human checks are still needed, especially for decisions that affect money or people.
• Privacy. Cameras can capture faces and number plates. Keep only what you need and restrict access.

Where footage includes people, look into your obligations under UU No. 27 Tahun 2022 tentang Pelindungan Data Pribadi (Indonesia's Personal Data Protection Law) and consult a legal professional.

Next steps

1. Pick one narrow, well-defined problem, such as counting vehicles at a single gate.
2. Decide how accurate the results need to be, and what happens when the system is wrong.
3. Collect and label data from the real camera.
4. Run a trial alongside your manual process for a few weeks.
5. Continue only if results hold up across different conditions.