Potestates/focus-frame-point-annotation

FOCUS — Foreign-Object Point Annotation

What you are doing

For every foreign object in a surgical frame, drop one point on it and pick its class.

You are not counting. We already have the counts. What the model is missing is where the objects are — it knows "this frame contains 5" but not which 5 pixels, which is exactly why it cannot learn to count.

Getting started (1 minute)

unzip shardN.zip -d shardN
cd shardN
python3 -m http.server 8000

Open http://localhost:8000 in Chrome or Firefox.

(Double-clicking index.html also works, but the local server is more reliable — some browsers disable local storage on file://, which would lose your progress.)

Controls

Every edit autosaves in your browser. Closing the tab loses nothing. When you finish, click Export JSON and send back the downloaded shardN_points.json.

The eight classes

1  Clip            surgical clip — small metal/plastic clip closing a duct or vessel
2  Sponge          gauze / swab
3  Needle          suture needle
4  Specimen        excised tissue
5  Specimen bag    retrieval bag
6  External drain  drainage tube
7  Silicone loop   silicone retraction loop / vessel loop
8  Gallstone       gallstone

Three rules that matter most

1. Instruments are NOT foreign objects

Graspers, scissors, hooks, suction devices, needle holders, trocars — these are surgical instruments held by the surgeon. Do not annotate them. A foreign object is something that stays in, or could be left behind in, the body.

Rule of thumb: attached to a shaft coming in from the edge of the frame → instrument, skip it. Sitting on tissue on its own, or clipped/sutured onto tissue → foreign object, annotate it.

This is the single easiest mistake to make, and it is exactly the error our model already makes — it confuses graspers with clips. Mislabelling here feeds the bug straight back in.

2. One point per object, at its centre

Three clips in a row = three points. One gauze, however large = one point. Half-occluded but still visible = annotate it. Completely hidden = do not annotate.

3. The reference counts are a check, not a quota

The panel shows e.g. Clips 2 / 3 — you placed 2, our label store says 3.

If you look carefully and only see 2, leave it at 2. We need real positions. Invented points to hit a number will actively teach the model the wrong thing. If a whole frame is unclear, skip it (leave it with no points) — we can detect those when we merge.

Pace

About 440 frames and ~1,870 points per shard, roughly 2 hours.

No need to rush. A wrong point hurts far more than a slow one — a point with the wrong class corrupts both "how many clips" and "how many classes" at the same time.

Shards

No video is split across two people — frames from the same video look very similar, so splitting one would bake in per-annotator differences as noise.