GREENPOINT / FROZEN PUBLIC-DATA CHECK / 03 NOV 2026

Evidence includes
what we missed.

Assisted review prototype. Weed-recognition target not met.
Three correct weed predictions do not establish reliable weed recognition. Low recall makes automatic treatment inappropriate.

20 held-out images. Every annotated plant counted.

We fixed five development images and twenty images from the official PhenoBench validation split before inference. We compared tight-crop baseline, same-source context, and context with a known development crop reference. The baseline won the predeclared development criterion: weed F1 minus twice the crop-as-weed rate, then precision and simplicity. Evaluation labels were used only to score the frozen pipeline.

Precision is based on only three positive predictions. No confidence interval or farm-level accuracy claim is implied. Missed and uncertain weeds remain in the recall denominator. Boxes are greedily matched one-to-one at intersection-over-union ≥0.5. Partial border instances are included. This is a small custom check, not the official PhenoBench benchmark or mask-quality leaderboard.

What the comparison changed

On development data, baseline and context-only predicted no weeds. The reference strategy predicted six weeds: two correct weeds and four crops. Its crop-as-weed rate was 4/40 (10%), with one incomplete classification response retained as a partial image. That is why references remain an optional experimental comparison and the baseline is the default. We did not tune the model using the twenty evaluation images.

Real moving video: integration, without a tracking claim

Two genuine four-second excerpts from the Bonn Sugar Beets 2016 robot-camera video ran through SAM native video. Both returned zero candidates. This is not evidence of weed-free ground or successful identity tracking. The source has per-frame semantic annotations, but its object keys are not reused across frames; it cannot provide an identity-tracking benchmark.

The recorded workspace offers the first evaluation frame, the first evaluation image with a model weed prediction, and the empty-candidate video result. These are labeled replays of actual Meta runs. Opening a page makes no paid model call. Human decisions are separate and initially pending. Source-image coordinates are not GPS.

Reproduce and inspect

Per-image predictions and independent reference boxes · Frozen manifest and source SHA-256 hashes · Machine-readable summary · Source and runners

Public-data inputs only. No measured farmer time saving, pesticide reduction, physical robot integration or geographic mapping. Native video explicitly sends the selected short clip to Meta; default sampling sends bounded snapshots. Server credentials are never included in public artifacts.

Data attribution

PhenoBench — Jan Weyler, Federico Magistri, Elias Marks, Yue Linn Chong, Matteo Sodano, Gianmarco Roggiolani, Nived Chebrolu, Cyrill Stachniss and Jens Behley. “PhenoBench — A Large Dataset and Benchmarks for Semantic Image Interpretation in the Agricultural Domain,” IEEE TPAMI 46(12), 9583–9594, 2024. Dataset and citation.

Sugar Beets 2016 — Nived Chebrolu, Philipp Lottes, Alexander Schaefer, Wera Winterhalter, Wolfram Burgard and Cyrill Stachniss. “Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields,” IJRR, 2017. DOI: 10.1177/0278364917720510. Original dataset. MP4 sample via pinned Supervisely sample.

Source data and the published JPEG/video excerpts, crops and derived benchmark annotations are distributed under CC BY-SA 4.0. Changes: JPEG encoding; four-second video excerpts resized to 1280×954, H.264, 10 fps; model overlays and derived bounding boxes. No endorsement by data authors is implied. Attribution details.