Instructions and covered skills
Read each scenario carefully before selecting an answer. Focus on the stated annotation policy and the visible image evidence. Avoid making assumptions that are not supported by the question. Work in a quiet setting and turn off notifications before beginning. Stay focused on bounding-box quality decisions rather than general computer vision concepts. Review your selected response when a scenario includes occlusion, cropping, or overlapping objects.
Key Areas
This test covers the quality-control practices used when producing bounding-box labels for object-detection datasets. Candidates interpret annotation policies that define what counts as an object instance, which visible pixels belong inside a box, and when an object should be omitted. The assessment addresses tight box placement, ensuring that boxes cover the intended object without unnecessary background or clipped object pixels.
Scenarios also examine partial visibility. Candidates must distinguish occlusion, where part of an object is hidden by another scene element, from truncation, where the image boundary cuts off part of the object. They apply visibility thresholds consistently and recognize when an annotation should be escalated because the policy does not resolve the case. Object separation is another focus, including touching objects, overlapping objects, grouped items, reflections, and displayed images.
Quality measurement is included through concepts such as intersection over union, duplicate annotations, missed instances, class mismatches, and systematic placement drift. Candidates should understand how reviewer findings can reveal policy gaps, annotator inconsistency, or tool-related errors. The test emphasizes decisions that can be reproduced across annotators and image batches.
Recommended Preparation
Review a written bounding-box guideline before taking the test, paying close attention to inclusion and exclusion rules. Practice drawing boxes around fully visible, occluded, truncated, and overlapping objects. Compare your labels with reference labels and note whether disagreement comes from object identity, box geometry, or interpretation of the policy.
Become familiar with intersection over union as a measure of geometric agreement, while remembering that a strong overlap score does not correct an incorrect class or a missed object. Practice reviewing samples in batches to spot repeated oversized boxes, repeated undersized boxes, and inconsistent treatment of similar scenes. Finally, develop a habit of documenting unresolved edge cases so that a project owner can clarify the policy and maintain consistent labels across the dataset.