Resources
Egocentric dataset datasheet and sample manifest
Connect episodes, streams, calibration, timestamps and permissions in a reviewable package without claiming missing evidence exists.
Published 2026-10-27 · 6 min read
| Record group | Suggested fields |
|---|---|
| Episode | episode_id, task, start/end convention, accepted duration |
| Stream | stream_id, camera role, source file, duration and checksum |
| Evidence | calibration_ref, timestamps_ref, quality_status |
| Governance | annotation_version, permission_ref, restrictions and package version |
A datasheet explains what a collection is and how it was made. A manifest maps that description to delivered files. Use both: a narrative hours claim cannot replace a resolvable file list, and a file list cannot explain permissions or intended coverage.
The downloadable CSV is a blank planning template, not a measured sample or evidence of Fiund inventory. Fill fields from actual records and explicitly mark unknown or unavailable values.
Use episode and stream identities
Give each distinct activity sequence an episode ID and each camera stream a stable stream ID. A row should identify its source file and camera role, such as head, left wrist or right wrist. Keep derivatives linked to their source IDs.
Record original duration, frame count where known, native image size and codec, plus checksum and transfer status. A manifest row should resolve to the actual version delivered, not a similarly named file left in another folder.
Reference geometry and time rather than repeating assumptions
Link each stream to applicable intrinsics and distortion files and each relevant pair to a transform version. State units, coordinate conventions and session applicability. For moving head-to-wrist relationships, distinguish tracked transforms from a fixed calibration.
Reference timestamp logs, clock domain, time units and synchronization evidence. Include any known gaps, corrections or resampling. A shared episode ID connects records logically; it does not prove simultaneous exposure.
Document labels and permissions separately
Identify annotation file paths, schema version, method and review status. Distinguish human review, sensor-derived output and model estimates. Link labels to the image or timeline version they describe so a corrected export does not silently invalidate them.
Use permission-status references rather than embedding identity documents or participant details in a broadly shared CSV. Establish an appropriate controlled review process for underlying records. The manifest is not a legal clearance certificate.
Run a package consistency check
Check uniqueness of IDs, required files, checksums, positive durations and expected stream membership. Reconcile accepted episode duration separately from stream duration. Review a sample with both ordinary records and missing-data cases.
Keep package version, preparation date, exclusions and unresolved findings in the datasheet. Agree a sample-sharing process first; do not publish private source paths or raw personal records to make a package look complete. Fiund confirms availability and scope after review.
Frequently asked questions
Does a completed CSV prove a dataset is cleared?
No. It maps evidence and statuses. Underlying records and the proposed use still need review.
Should unavailable signals be left looking complete?
No. Use an explicit missing or unavailable status and explain its implications.