Identify what becomes hidden
Record whether the missing evidence is a hand, an active object, a tool tip, or the final task result. An object obscured by the participant is different from the entire lens being covered. This distinction helps explain whether an example is challenging scene content or a capture failure that should be corrected before additional recording.
Consider timing as well as duration
A short occlusion during a decisive contact can matter more than a longer occlusion during an idle interval. Review the hidden period against the task timeline. For event recognition, surrounding context may be acceptable; for contact localization, it may not. Define the policy around observable requirements rather than an arbitrary allowance for time out of view.
Make annotation uncertainty visible
Specify whether tracks continue through occlusion and whether labels may be inferred from later evidence. Keep visible observations separate from inferred events. An annotator who watches a full sequence may know more than an online model will know at the same instant, so the allowed context should reflect the intended training and evaluation task.
Use the pilot to set boundaries
Select examples with objects behind hands, activity below the frame, and interactions near clutter. Have reviewers apply the proposed rules and note disagreements. Decide which cases to retain as deliberate difficulty and which fail the brief. Share those decisions with Fiund when discussing capture feasibility, since viewpoint changes may improve one interaction while obscuring another.
Review checklist
- Identify the hidden evidence
- Review occlusion at critical events
- Separate inference from observation
- Label intentional difficulty and capture failure