Image Review: Spot People in Security Camera Footage

Job description We are building a training dataset for a computer vision system and need a small team to review machine-generated results. The work is simple and repetitive. You will see a grid of images on a web page. Some show a small crop where our system thinks it detected a person. Others show a full camera view that is supposed to be empty. You click the ones that contain a person, then submit, and the next grid loads. There is nothing to draw, nothing to type, and no software to install. It runs in a normal browser. Most people find a grid takes under a minute once they are into it. We will send you a one-page guide before you start. It covers the edge cases: people on TV screens, reflections in glass, mannequins, jackets over chairs, and so on. It takes five minutes to read and it answers almost everything. What we are looking for Careful attention over speed. We are paying hourly precisely so nobody feels pushed to rush. Comfortable with repetitive visual work. It is not difficult, but it is sustained. Reliable availability across the days below. We would rather have six people who finish than twelve who start. Willingness to flag things. If you see something the guide does not cover, tell us. Those messages genuinely improve the work. No prior annotation experience needed. No particular technical background needed. Timing Starts Sunday. We will send the guide and your login on Saturday so you can read it in advance. Roughly 10-15 hours each, over Sunday and Monday. Flexible within those days. Work when suits you, in whatever blocks you like. Some flexibility on total hours depending on how quickly the team gets through the queue. Pay We will negotiate a fair hourly rate, tracked through Upwork. Breaks are paid. Accuracy matters more to us than throughput, and tired reviewing is worse than no reviewing. How the work is checked Items with known answers are mixed into everyone's queue throughout the job. They look identical to normal items, and they run from the first screen to the last. This is standard practice for dataset work. It is how we confirm the finished data is trustworthy. It also means consistency matters more than any single decision: nobody is judged on getting one hard image right, but sustained accuracy across the whole job is what we are paying for. If accuracy drops well below the standard and stays there, we will end the contract early. You are paid for all tracked hours regardless. About the material The footage is from a public research dataset of ordinary indoor and outdoor scenes. No sensitive or graphic content. Nothing you see is confidential and nothing needs to be kept private. We are hiring 6-10 people and moving quickly, so short and specific beats long and polished. We will confirm within a few hours of applying.