Boxes, polygons, or masks: pick by the question you are asking

The most expensive labeling mistake is not bad labels. It is the right labels for the wrong question. Teams default to the richest annotation type they can afford, then discover the model they actually needed was answerable with boxes at a fraction of the cost.

Is there a vehicle in the landing zone? How many people crossed the fence line? Bounding boxes answer presence, count, and rough location, and they are the cheapest label on the menu. For most monitoring and alerting work, that is the entire question, and boxes are the right call at any volume.

Measuring the area of a rust patch, the footprint of standing water, the exact edge of a crop row: these need the shape, not just the region. Polylines handle linear assets like pipelines and power lines. You pay more per label because the answer carries more information.

Full segmentation, the most expensive label of the three, is for questions where boundaries are the product: medical imagery, precision agriculture, surface defect mapping. If you find yourself measuring things off the mask, it was worth it. If you are just checking presence, you bought a scalpel to open mail.

Start with boxes. Upgrade the classes that prove they need more. Your first model will tell you which questions it cannot answer, and that list is your polygon budget.