What we sell is not data but rights
The main public sets such as EgoDex and Ego4D carry non-commercial licences, so they cannot be used for commercial training. What Aeffer sells is data for which demonstrators have consented to commercial training.
Three forms of use
| Tier | Scope of use | Redistribution | Commercial training | Notes |
|---|---|---|---|---|
| Evaluation | Internal benchmarks · reproducing papers | No | No | Sample segments only |
| Commercial training | Training your own models · shipping in products | No | Yes | The standard contract |
| Training without export | Training inside our infrastructure · only weights leave | N/A | Yes | The data never leaves · from hub stage H1 |
What you will ask first
- Can we use it for commercial training?
- Yes. That is the reason this data exists. The main public sets — EgoDex and Ego4D (egocentric video) and AMASS (motion capture) — carry non-commercial licences, so they cannot be used for commercial training. Aeffer only distributes data for which demonstrators have consented to commercial training.
- Who owns the weights of a model trained on it?
- You do. We claim no rights over the weights or anything derived from them. We restrict redistribution of the data itself, nothing more.
- Can we redistribute it?
- Not under the base licence. You may not resell the data or upload it to a public repository. If you need redistribution rights, we negotiate them in a separate contract.
- What happens if a demonstrator withdraws?
- Their session is removed from versions distributed after the withdrawal. Copies already downloaded carry a contractual obligation to destroy them, but weights already trained cannot be recovered. We state this explicitly in the contract.
- Is there copy protection (DRM)?
- No. Technical copy protection does not work on training data — if a model can read it, it can be copied. Instead we manage it three ways: contract, per-customer fingerprinting, and training without export (compute-to-data).
- How is personal data handled?
- Under the Personal Information Protection Act we take separate consent for third-party provision and for overseas transfer. When providing to an overseas buyer we disclose the destination country and purpose to the demonstrator, and only consented data leaves.
We apply no technical copy protection
DRM cannot work on training data in principle. If a model can read it, it can be copied, and once data has been absorbed into weights there is no way to retrieve it. Applying it would only inconvenience the buyer without protecting anything.
So we manage it three ways.
- Contract — we set out the ban on redistribution and the duty to destroy, and define liability for breach.
- Per-customer fingerprinting — each distributed copy carries identifiable micro-variations, so a leak can be traced.
- Training without export — training runs on our infrastructure and only the weights leave; the data does not. This opens at hub stage H1.
In Korea, the misappropriation-of-data provisions of the Unfair Competition Prevention Act and database producers' rights apply on top of this.
We check once more before anything leaves the country
Physical demonstration data is personal data. Under the Personal Information Protection Act we take separate consent for third-party provision and for overseas transfer. If the buyer is a foreign entity, we disclose the destination country and purpose to the demonstrator, and only data consented to that scope is sent.
We do not provide beyond the scope of consent. However large the contract, we do not sell data we do not have consent for. If that principle bends, the whole business collapses.
The standard contract template will be published when the first pilot is signed. It will be finalised after legal review; what is on this page is a summary of the contract terms, not the contract itself.