How to Brief a Computer Vision Project: Counting, Quality Inspection and Camera Monitoring

Computer vision makes a camera understand what it sees: counting boxes on a production line, spotting defective products, reading numbers and text, tracking movement and monitoring quality. Success in these projects depends more on real details at your site than on the name of the model used. This guide helps you write a clear post in the AI & Machine Learning category on Farasa. For the general rules, read our guide to posting software projects.

Write what you want from the camera: counting, defect detection, sorting types, reading text or numbers, or tracking a person or vehicle. And write the decision it feeds: “stop the line if defects exceed a threshold,” or “calculate daily output automatically.” The business goal sets the accuracy you need and how the system runs.

This is where projects differ most. Describe what the camera sees: what objects, how big, how many per frame, whether they move and how fast, whether they overlap or pile up, what the lighting is like (steady or changing, day and night), and whether there is dust, glare or shadow. Attach real photos or video from the site, because a model that works on clean pictures can fail in your factory.

Mention the cameras you have (type, mounting angle, resolution) or say you need a recommendation. Is analysis real-time or on recordings? Where will it run: a PC on site, a small device next to the camera, or a remote server? This decides the processing speed needed and the hardware cost, and the internet at your site may not be reliable.

A model learns from labelled examples, meaning images in which someone has marked the objects or defects. Say how many images you have, whether they are labelled, and who will label the rest. In defect projects the faulty images are usually rare, so say so plainly because it affects the method and cost. The plan may include a week of collecting images at your site before building the model.

Don’t ask for “100% accuracy”; it isn’t realistic. Write a measure that fits your business: “count boxes with under 2% error in normal lighting,” or “catch 95% of defects with a limited number of false alarms.” Think about which mistake is worse: missing a defect, or stopping the line for nothing? That guides how the system is tuned.

Before investing in every line, agree a pilot on one line or one camera to measure real accuracy and find the site’s problems, then expand. Decide what will be delivered: the model, the runtime code, a way to retrain when the product changes, and installation documentation.

If the camera will capture people, say so, check the rules and consents in your organisation, and agree how images are stored and deleted. Monitoring staff needs particular care.

“Counting cans on a conveyor belt at a factory in Erbil. The cans are cylindrical with similar diameters, pass at a moderate speed and sometimes touch. One camera is mounted above the belt, and the lighting is steady except for daylight near a window. I have 3 hours of video and 300 unlabelled images. I want a daily count with under 2% error, running on a small computer on site without internet, with a screen showing the count. First phase: a pilot on one line. Budget: fixed price for the pilot within a range (put your range here). Start your proposal by naming a computer vision project you’ve delivered in an industrial setting.”

Write the task and the business decision, describe the real scene with photos, state the cameras and hardware, set a realistic accuracy measure, and start with a small pilot. That gets you proposals you can compare. Post your computer vision project on Farasa.