Resources
Off-the-shelf versus custom egocentric collection
Choose an existing subset, new recording or a hybrid only after reviewing task coverage, missing signals and pilot cost.
Published 2026-10-27 · 5 min read
| Question | Evidence or decision |
|---|---|
| Existing data | Does a reviewed, permitted subset meet every essential requirement? |
| Custom | Can a pilot demonstrate the missing requirement feasibly? |
| Hybrid | Can the added coverage share schema and source-aware split rules? |
| Cost | Compare complete accepted packages and remaining uncertainty |
Choose an existing dataset when a verified subset fits the task and licence. Choose custom collection when an important requirement cannot be met from existing material and the pilot supports a feasible scope. A hybrid can reuse suitable data and commission only the missing coverage.
Off-the-shelf does not mean immediately cleared for every use; custom does not guarantee acceptance or model performance. Compare the actual proposal rather than applying a universal faster-versus-better rule.
Start with a task-to-evidence matrix
List target actions, environments, operators, sequence structure, views, geometry, timing and labels. Mark each as essential, optional or unnecessary for this task. This prevents buying extra channels while overlooking a missing critical signal.
For an existing offer, inspect evidence against each essential requirement. Missing recorded sensors cannot be recovered merely by changing a file extension. Derived depth or pose may be useful, but should be evaluated as a different deliverable from native measurements.
Ask whether the existing subset really exists
Request representative originals, manifests and source documentation. Distinguish verified in-scope material from a provider’s proposed sourcing capacity. Check duplicate episodes, incomplete views and actual accepted activity rather than relying on the catalog headline.
Identify preparation still required: extraction, labels, redaction or timing review. Include it in the quote and schedule. An existing file library may need meaningful work before it is a usable licensed package.
Use custom collection for a documented gap
Specify the missing requirement and confirm that site access, operators and equipment can actually support it. Run a pilot before committing to large recording volume. Review task coverage, technical evidence, permissions and package loading.
For example, an existing single-view household subset might suit an observation task but not a requirement for synchronized close interaction views. That is an illustrative decision, not a claim of real Fiund inventory or a validated robotics training outcome.
Compare total accepted scope
Price the existing licence and preparation against custom setup, capture, review, rework and licence scope. Include uncertainties and decision gates. Do not assume that new recording will be accepted at the same rate as a selected existing sample.
A hybrid needs compatible schemas and clear source-group boundaries; do not mix related samples across evaluation splits. Fiund can review the requirement and proposed routes, but confirms actual supply and commercial scope only after assessment.
Frequently asked questions
Does custom collection guarantee a useful dataset?
No. The pilot and acceptance process must establish fit.
Can annotations turn mono video into native stereo capture?
No. Derived representations should be identified and evaluated as such.