Distinguish the hidden cases
Occlusion by another object differs from truncation at the image edge. A target can also be visible only in a reflection or through transparent material. List the cases your task includes and give examples. A single partial-visibility flag may be too broad if these conditions trigger different annotation or evaluation decisions in the project.
Define the evidence threshold
Specify when enough evidence remains to assign a class and when reviewers should mark uncertainty. Avoid requiring an exact identity that depends on unseen information. Decide whether a fragment receives a broad class, a special flag, or no annotation. Keep these decisions consistent between ordinary images and deliberately difficult examples in the same collection.
Align geometry with policy
Boxes and masks need separate instructions for occluded objects. A box might describe visible extent while a mask follows disconnected visible regions. If estimated hidden extent is required, identify it as a different annotation task and review its assumptions explicitly. Do not allow annotators to switch conventions depending on what seems convenient in an individual image.
Review crowded examples together
Use a pilot containing overlapping objects, partial groups, and small fragments. Ask reviewers to explain disagreements so the policy can be refined before scaling. Track occlusion attributes in evaluation reports to reveal whether a failure is concentrated in a specific condition. Source availability and the feasibility of consistent annotations should be reviewed for the particular collection.
Review checklist
- Separate occlusion from image-edge truncation.
- Define the minimum evidence for a class label.
- Choose visible or estimated geometry explicitly.
- Pilot crowded scenes with reviewer discussion.