docs: name the backend axis correctly and precise the v0.9 blocker - #11
Merged
Conversation
Two things went stale and were found by running the pipeline on real data for the first time. "SAM vs. Cellpose" is not the axis. CellposeSegmenter runs Cellpose-SAM v4 (Pachitariu et al. 2025), which is itself SAM-backboned — the two learned backends share an architecture and differ in what they were trained on. The real axis is zero-shot generic vs. domain-finetuned, and the companion plan's H_A already states it that way: a specialist fine-tuned on <= 300 hand-labelled stones against zero-shot SAM. The paper's literature guide says the same thing from the other side, calling ImageGrains "Cellpose for stones". v0.9's blocker is annotation, not capture. A real orthophoto and DEM exist (paper/input/data/Export/, Tiberias 2023-03-10) and the pipeline has now been run on them end-to-end. What is missing is ground truth — no reference stone outlines anywhere in the paper repo — so no metric can be scored. H_A bounds the work at <= 300 instances, which makes it a schedulable task rather than a wait for a season. Add the freeze-first task that follows from this: looking at model output on a scene before its hold-out status is designated spends the scene. v1.0's blocker is restated as the v0.9 verdict, since the capture it named now exists. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Documentation only. Both corrections were found by running the pipeline on real excavation data for the first time.
"SAM vs. Cellpose" is not the axis
CellposeSegmenterruns Cellpose-SAM v4 (Pachitariu et al. 2025,Pachitariu_2025in the paper's bibliography), which is itself SAM-backboned. So the two learned backends share an architecture and differ in what they were trained on — the roadmap's phrasing implied two families.The real axis is zero-shot generic vs. domain-finetuned, which is what the companion plan's H_A already says: a specialist fine-tuned on ≤ 300 hand-labelled stones against zero-shot SAM. The paper's own literature guide reaches it from the other side, calling ImageGrains (Mair et al. 2023/2026) "Cellpose for stones".
v0.9 is blocked on annotation, not capture
The milestone said "needs real trench data". The raster exists —
paper/input/data/Export/, Tiberias 2023-03-10 — and the pipeline has now been run on it end to end. What does not exist is ground truth: no reference stone outlines anywhere in the paper repo, so no metric can be scored.That distinction matters because it changes the shape of the work. H_A bounds it at ≤ 300 hand-labelled instances, which is schedulable, rather than a wait for a field season.
A task is added ahead of the metrics work: freeze the hold-out scene IDs first. Looking at model output on a scene before its status is designated spends that scene, and plan Task 1.1 is still open.
v1.0's blocker is restated as the v0.9 verdict, since the capture it named now exists.
Not in scope
The corresponding line in
paper/input/ideas/structura-plan.md— "no research data (orthophotos/DEMs) or code exist yet" — is also stale, butinput/is human-owned perpaper/CLAUDE.mdand the plan is RE-FROZEN. The finding, a proposed replacement sentence and the contamination question are recorded in the paper'sknowledge/_meta/log.mdfor sign-off instead.Verification
ruffclean,mypyclean on 26 files. Tests pass (a pre-existingSamSegmenterbug surfaced separately once thesamextra was installed — CI has never exercised that path; fix follows in its own PR).🤖 Generated with Claude Code