labeler

The labelers — the web tools the labeling team uses to produce Cornerman's training labels. One folder per labeler; every tool saves straight to the Box Labeled Data Sheet via the shared Apps Script backend (no login).

Start here: https://cornerman-ai.github.io/labeler/ — the landing page lists every labeler. First-time Sheet/Apps-Script setup: SETUP.md.

The labelers

Folder	What it labels	Why it exists
punch/	Punches in a round: start/end time + type per hand	The main tool. Its labels are the training corpus for the punch classifier
unusable/	Stretches of a video where the BlazePose skeleton can't be used (with a reason), a reviewed mark per video, and the whole-video retirement into Skeleton Problems	The unusable-footage pass, made AFTER extraction (the skeleton doesn't exist while rounds and punches are labeled); feeds the roll detector's exclusion mask and the skeleton-usability model
impact/	The exact impact frame of each labeled punch	Trained the impact-frame spotter
punch_directions/	16-way (22.5°) direction of straight punches	Trained the axiality (punch-direction) model that three rules gate on
bodyshot/	Head vs body, one short clip per punch	Refines punch-type labels where head/body was ambiguous
callout/	Called-out combos in guided rounds vs what was actually thrown	Measures how well users follow the app's callouts
rules/	Form-rule spot checks on rounds	Ground truth for rules-engine tuning
chin_tuck_1.0/	Three chin-vs-shoulder questions per sampled frame	Current chin labeling. Repeated samples measure each labeler's self-agreement ceiling
chin_tuck_0.0/	One verdict (tucked/level/air) per sampled frame	The original chin labeler, frozen for reference. Has no data of its own — reads 1.0's chin_frames.json/frames/ by relative path
uppercut_drop/	How far the punching hand drops before each uppercut: Great / Good / Bad / Very bad / Can't tell, one looping clip per labeled uppercut, ~10 % blind repeats	The coach's ratings set the cut-offs of the "keep your hand up before the uppercut" score
roll_eyes/	Do the eyes stay up through each roll: Good / Mid / Bad / Can't tell (with a reason) + a comment, one looping clip per labeled roll, ~10 % blind repeats; the name admin shows the model's head angle and every rater's answers	The coach's ratings set the cut-off of the "keep your eyes up when you roll" rule
roll_legs/	Does the boxer use the legs to get under the punch in each roll — Uses the legs / Doesn't / Can't tell (with a reason) + a comment, on the roll eyes clips; the scale is one PASS block in roll_legs.js, an example per rating, ~10 % blind repeats; admin shows the skeleton's knee bend and lean	The coach's ratings set the cut-off on the knee bend
roll_shoulders/	Do the shoulders rotate through the roll — Good / Bad / Can't tell (with a reason) + a comment, on the roll eyes clips, frontal rolls only; the scale is one PASS block in roll_shoulders.js, an example per rating, ~10 % blind repeats; admin shows the skeleton's shoulder turn	The coach's ratings set the cut-off on the shoulder turn
bag_bite/	Does each punch on the heavy bag snap or push — two buttons, Snapping / Pushing, + a comment, one looping clip per punch cropped to the boxer, the skeleton on it and beside it with the facing angle; 200 punches, each once; admin shows every rater's answers	The coach's ratings say what rule 41 (punches stuck in the bag) should measure
guard_drop/	Hands-down moments in a round	Ground truth for the guard-drop rule
hip_rotation/	Rotation-quality rubric per punch	Ground truth behind the hip-rotation rule
bladedness/	Pairwise "which stance is more squared?" (shoulder pairs + hip pairs), plus the coach review where the coach scores curated frontal frames	The too-squared/too-bladed research — pair judgements calibrate the candidate metrics; the coach review is the expert reference
axiality/	Review of ground truth vs model direction predictions	QA pass on the axiality model

Identity — one name, every tool

Every labeler shows the same slim top bar — "👤 Your name:" until filled, then "👤 Name · change" (shared/labeler_name.js). Pages that had their own name field (chin tuck, bladedness pairs) get it hidden and auto-synced; the bar is the one UI everywhere. The name is sent as the labeler field on every save — the Apps Script routes rows/tabs by it, unchanged. No more per-person URLs; an old ?labeler= link still works as a one-time seed. Pages with their own name inputs (chin tuck, bladedness pairs, orientation) stay in sync with the same stored name automatically.

Shared pieces

• shared/player.js — the video player + Sheet-posting core every video labeler uses
• shared/ui.js — the chrome every video page shares (transport icons, volume, speed menu, timeline scroll-zoom and ⌘ keys, picture zoom, folds, click-to-type time, status chips, name field, shortcuts sheet); punch/ui.js keeps only the punch page's own (strip dragging, round spans, context menu)
• shared/labeler_name.js — the ask-once labeler identity (see above)
• shared/style.css — common styling
• shared/videos.json — which videos the video-based labelers list (regenerate with shared/build_videos_json.py)
• apps_script/Code.js — the Google Apps Script backend (deployed as a Web App; auto-deploys from master via GitHub Actions — see apps_script/README.md); labels land in the Sheet at Drive Cornerman/data/labels/labeling_team/

Conventions

• A labeler's page, JS, data, and media all live in its folder — tools are self-contained; only truly shared code sits in shared/.
• Frame/clip media that pages display is exported by the backend research scripts (cornerman-backend/ml/research/<topic>/) and committed here so GitHub Pages hosts it.

Retired labelers live in cornerman-archive/legacy-labelers/ (orientation, 2026-08-07 — its 8-bin labels shipped as the ankle-orientation constants).