Read from LIT-385 — ARXIV-2312.14132.
The constraint it declines
The paper is explicit about what it is refusing, and the sentence is worth keeping whole: camera intrinsics and extrinsics are “tedious and cumbersome to obtain, yet they are mandatory to triangulate corresponding pixels in 3D space, which is the core of all best performing MVS algorithms.”
That is a description of a dependency order, not of a difficulty. Every classical method needs calibration and pose first because triangulation is defined in terms of them. So the interesting question is not “can we estimate poses better” but “is triangulation the only way to get 3D”.
What replaces it
Pairwise reconstruction as pointmap regression, which “relax[es] the hard constraints of usual projective camera models”. The projective model stops being something the method must satisfy on the way to an answer and becomes a property the answer turns out to have.
Two consequences follow immediately, and both are the kind of thing that signals a reframing rather than a better solver:
- Monocular and binocular unify. In the classical framing they cannot — one view has no baseline, so there is nothing to triangulate. Here they are the same problem with different amounts of evidence.
- The pipeline runs backwards. From the 3D output you “seamlessly recover pixel matches, relative and absolute camera”. Everything the old pipeline demanded as input is available as output.
More than two images need a global alignment step to put all pairwise pointmaps in one frame — the one place an explicit optimisation survives, and the thing LIT-384 later removes.
What I am not filing from this
The SoTA claims on monocular/multi-view depth and relative pose. They are the evidence that the reframing works, not instructions — and a practice that said “use DUSt3R” would date the moment the next model landed, which it did.
The practice records the dependency inversion, because that is what survives the specific network.
Why this matters to a record that is mostly about language models
Because the shape is general and the record has met it before. A pipeline whose stages exist because each was separately solvable gets replaced by one model that does the whole thing, and the intermediate quantities turn out to be recoverable rather than required.
That is the same move as SOTA-227’s draft-and-verify replacing hand-tuned decoding heuristics, and the same as end-to-end learned tokenizer-free models against staged NLP pipelines. The geometry instance is unusually clean because the discarded stage — triangulation — had a closed-form justification, and it still lost.