Sun et al. (2022) — ARXIV-2205.12257
Key takeaways
- Estimates object pose with no CAD model and — the point the previous takeaways inverted — no instance- or category-specific network training. The paper’s own words: it “can handle objects in arbitrary categories without instance- or category-specific network training”
- The recipe separates instance knowledge from learned knowledge: build an SfM point cloud of the object once from a scanned video, then match 2D interest points in a query image to its 3D points using a generic graph attention network that never saw that object or its category
- Adding a new object therefore means scanning it, not training — “one-shot” in the sense of one construction, not one example
Standing in the anthology
Read — NOTE-044. Nothing is sourced here.
The takeaway “category-level pose estimation” was precisely backwards.
Category-level estimation is the prior paradigm this paper positions itself
against. A reader taking that bullet at face value would attribute to OnePose
the approach it exists to avoid — the third bullet in this pass to state the
inverse of a paper’s claim, after LIT-025’s LayerNorm direction and LIT-108’s
dynamic scenes.
The division of labour is the transferable part: instance knowledge in a
structure you can rebuild without training, learned knowledge generic and
fixed — the same split as LIT-060’s frozen retriever with a swappable
database.
Audited by #121; unsourced, and the audit’s judgement is that it should stay so. R1 — put instance-specific knowledge in a structure you can rebuild without training and keep the learned component general — is a real claim, but it is the same decision SOTA-205 already files from four sources, arrived at for pose rather than for rendering. R2 is a vocabulary point about what “one-shot” means, which is worth reading and is not a practice.