Data: expand Sharoni-lab labels via model-assisted annotation instead of fresh labelling

[timlandgraf]

What

Get more labelled dances from the Sharoni lab (Avi Gabai), but not by having them annotate from scratch. Instead: run the current model over the unlabelled footage, load the predictions into the viewer dashboard, and have the annotator correct them.

Why

Annotating waggle runs from a blank slate is slow and is the reason our labelled set has grown so little. If the model is already reasonably good — and from what we have seen it is — then reviewing and fixing predictions is a much cheaper interaction than producing them, and it scales with model quality instead of fighting it.

The dashboard already has the pieces: GT annotation editing, per-video inspection, and re-clustering with adjustable post-processing.

Depends on

Model quality on Sharoni footage specifically. If precision is poor there, correcting predictions is more work than annotating fresh, and this approach backfires. Establish that first — a per-lab recall/precision breakdown at the current checkpoint is the go/no-go.

Note that the Sharoni 30-frame annotation offset (earlier milestone #14) has never been conclusively resolved. If existing Sharoni GT is misaligned, that has to be settled before we ask for more labels against the same convention.

Done when

• Per-lab detection quality measured on the current checkpoint; Sharoni numbers good enough to make correction cheaper than annotation
• Sharoni 30-frame offset question closed out
• Model-assisted annotation flow tried end to end on a handful of videos
• Proposal sent to Avi with a realistic estimate of effort per video

[timlandgraf]

Inventory check: what Sharoni footage we actually hold

Counted against fps_multires_clean.csv (654 video files, 160 base recordings) and the video store on horus:

Lab	Base recordings	Video files in annotations
berlin (0*)	96	288
nieh (T*)	62	360
sharoni (C*)	2	6

Both Sharoni recordings are 2-minute clips at 1224×1024, with derived fps tiers:

• C1_11_03_25_my_video-1-01.00.00.815-01.02.00.957 — val split, 27 labelled runs
• Capture_2025-04-29_14-59-11-1-00.00.01.000-00.02.00.688 — train split, 119 labelled runs

(146 labelled waggle runs in total. Two further spatial tiers per recording, 612×512 and 306×256, exist on disk but are not in the annotation set.)

Three consequences for this issue

1. There is no unlabelled Sharoni footage to run the model over. The 2148 "external videos" available to the dashboard (/mnt/horus/bee_dance_videos_tim) are the 2008 640×480 set and 2015 BeesBook material at various fps tiers — nothing from Sharoni. So model-assisted annotation cannot start from data we already hold: the first ask to Avi is raw video, not labels. That is a much smaller ask and worth making straight away, independent of the labelling question.

2. The go/no-go test proposed in this issue is not currently measurable. "Per-lab detection quality on Sharoni" rests on a validation set of exactly one 2-minute clip — the other recording is in training. Any Sharoni precision/recall figure we quote is a single-clip number. Getting a second and third Sharoni recording is therefore a precondition for the decision, not a follow-on from it.

3. This also bears on Decision: bee_size_multiplier stays a single global N (not per-lab) #1: BEE_LENGTH_FRACTION["sharoni"] = 0.0330 is derived from one video (n_source_videos = 1 in projected_summary.csv), and Sharoni's entire contribution to training and evaluation is two clips. A per-lab N for sharoni would be fitted against two clips — a further reason the multiplier stays global.

Revised first step

Ask Avi for raw recordings before asking for any annotation effort. Even a handful of additional 2-minute clips across different days would let us (a) measure Sharoni performance on more than one clip, (b) widen the comb-cell measurement base behind the bee-length constant, and (c) test whether correction-based labelling is actually cheaper than fresh annotation on their footage.