Source
Sun et al. (2022), LIT-111 — ARXIV-2205.12257.
The rule
Two kinds of knowledge get conflated whenever a model is asked to handle a specific thing: what is true of this instance and what is true in general. Factor them.
The instance’s structure is built once by reconstruction — in the source’s case an SfM point cloud from a scanned video. The general capability is a network that matches 2D interest points to 3D points and “can handle objects in arbitrary categories without instance- or category-specific network training”.
The operational consequence is the point of the practice. Supporting a new object becomes a scan rather than a training run: no labels for it, no fine-tune, no risk to anything the network already does. “One-shot” here means one construction, not one example.
The factoring originates in LIT-111, OnePose, which estimates object pose with no CAD model and no instance- or category-specific training: everything about the object lives in an SfM point cloud built from one scanned video, and a generically trained graph attention network matches 2D interest points in the query image against it. Its central demonstration is that this works on objects of arbitrary categories, which is what makes “scan, don’t train” a claim rather than a hope.
When it applies
- When the instance has stable structure that can be reconstructed. A rigid object does; a deformable or featureless one does not
- When the set of instances grows after deployment. If the instance set is fixed and known at training time, the factoring buys nothing
- When retraining is the expensive part — because a team cannot retrain, or because retraining risks regressions elsewhere
Why it is filed outside its domain’s frame
The source is an object-pose paper and the recommendation is not about pose. It is about where instance knowledge lives, and the same question is asked whenever a system is expected to know about particular entities: retrieval against an index instead of memorisation in weights is the same factoring, reached from a different direction.
Filed under ADR-026’s rule — the claim takes its kind rather than the domain
it was discovered in — and model-architecture is its second tag for that
reason.
Conditions
- One paper, no replication in this record. See
consensus_note - Reconstruction quality bounds it. The matcher can only be as good as the point cloud, and the paper’s setting is a scanned video under reasonable conditions
- The generic matcher is doing real work, and is the part that would need re-training if the class of instances changed rather than the instance