Image labeling is the second-busiest task category on trAIn, and one of the most accessible ways into AI training work — no writing required, just careful eyes. But it has a rejection pattern that catches nearly every beginner: the mistakes are visual, subtle, and consistent.
Here's how the work actually gets reviewed, and how to stay on the right side of it.
The two main flavors
Bounding boxes ask you to draw a rectangle around every instance of a target object — every car, every pedestrian, every product on a shelf. Fast work, but the precision bar is higher than beginners expect.
Segmentation asks you to trace the object's actual outline, pixel-close. Slower, better paid, and much less forgiving of sloppy edges.
Both get reviewed the same way: against the campaign's written rules, against golden (pre-labeled) images seeded into your queue, and against what other trainers labeled on the same image.
The five mistakes that get work rejected
1. Loose boxes. A bounding box should touch the outermost pixels of the object — not float around it with a margin, and not clip its edges. "Close enough" is the single most common rejection reason. Zoom in. Always.
2. Missing instances. You boxed four of the five cars and missed the half-visible one at the edge. Reviewers count omissions, and clients notice them most. Before submitting, do one deliberate sweep of the whole image asking "is there anything else?"
3. Labeling occluded objects wrong. Half an object behind a pole: do you box the whole estimated shape or only the visible part? Campaigns split on this — some want amodal (estimated full) boxes, some want visible-only. The instructions will say. This one rule causes more rejections than any other, and it's a two-minute read.
4. Over-labeling. Boxing things that almost match the target — a van when the campaign wants cars, a decorative image of a dog when it wants real animals. When unsure, the answer is in the examples section of the guidelines, not in your gut.
5. Rushing segmentation corners. On segmentation tasks, beginners lose accuracy at object corners and thin structures (chair legs, bicycle spokes, hair). Slow down exactly there. Straight edges are free points; corners are where reviews are won and lost.
Golden images are your friend
Every campaign mixes in pre-labeled "golden" images to score you. You can't tell them apart from real work — that's the point. But it means your accuracy score is always being measured, on every image, not just when someone manually reviews you. There is no "nobody will notice this one."
The upside: golden images make feedback immediate. If your accuracy dips, you'll see it on your dashboard before rejections pile up — treat that dip as a signal to re-read the guidelines, not as bad luck.
Getting started
If you're new, begin with bounding-box campaigns before touching segmentation — the skill transfers, and the reviews are more forgiving while you build speed. Our beginner's guide to image annotation covers the tooling basics, and when your accuracy is solid, the image-focused certifications are your route into the better-paying queues.
Careful labelers are rare, and clients pay for rare. Be boring, be precise, and the approvals will take care of themselves.