Birkl et al. (2023) — ARXIV-2307.14460
Key takeaways
- A controlled backbone sweep, not a method. The MiDaS architecture and recipe are held fixed while the image encoder varies across BEiT, Swin, SwinV2, Next-ViT, LeViT and recent convolutional approaches — where v3.0 used vanilla ViT alone
- The design answers a question the backbone papers cannot: does classification performance transfer to depth estimation? Including convolutional approaches “that achieve comparable quality to vision transformers in image classification tasks” is the direct test
- 28% depth-quality improvement from the best backbone; the deliverable is a performance-runtime frontier, with efficient models “enabling downstream tasks requiring high frame rates” — a zoo, not a winner
- Also publishes the general process for integrating a new backbone, which is what keeps a model zoo alive after the paper
Standing in the anthology
Read — NOTE-047. Nothing is sourced here.
The methodological sibling of LIT-156 (Fantastic Pretraining Optimizers),
read in batch B: both hold a setting fixed and sweep one component fairly, and
both find that a component’s reputation elsewhere does not settle its performance
here. Two independent instances of the same lesson — a component that wins on
one task is not thereby the right choice for another, and only a fair sweep
tells you — and the general form is stated nowhere in the record.
The takeaways it carried described the MiDaS line rather than this release, whose entire content is the comparison.
Audited by #121; unsourced. R1 — report a frontier rather than
a winner when downstream uses have different cost constraints — is general,
analysis-and-evaluation, and the reading notes R2 is the same lesson
LIT-156 teaches. Two in-record sources for a reporting norm is the
threshold this audit used elsewhere, and the reason it is not filed here is
narrower: the record’s evaluation practices are about what to measure, and this
is about how to present it. That may be a gap in the topic rather than in the
corpus.
Carries no practice, and #154’s pass says why. The transferable claim
here — classification accuracy does not by itself pick a backbone for dense
prediction, so sweep it — is real and the paper measures it (28% depth
quality from the best backbone, and the deliverable is a performance-runtime
frontier rather than a winner). But the record’s instrument for that claim is
a benchmark note, not a recommendation: what a practitioner does with it is
run their own sweep, which the paper already tells them and this record cannot
improve on.
What it supplies instead is the process for integrating a new backbone,
which is the part that keeps a model zoo alive after publication and is
infrastructure rather than advice (ADR-032).