kodokan

Study Kodokan Judo throws from video via body-pose analysis: download technique demonstrations, extract two-person (tori/uke) skeletons, split each clip into its repeated demonstrations, visualize them, and compare/score demonstrations.

from kodokan.acquire import download_techniques
from kodokan.track import estimate_poses_tracked
from kodokan.segment import segment_demonstrations
from kodokan.viz import render_skeleton_video

res = download_techniques(playlist_items="2")[0]          # Seoi-nage (#002), with metadata
seq = estimate_poses_tracked(res.path, source_url=res.info["webpage_url"])  # tracked tori/uke COCO-17
demos = segment_demonstrations(seq, min_two_person_frac=0.3)                # per-demo (start_s, end_s)
render_skeleton_video(seq, out_path="overlay.mp4", source_video=res.path)  # skeletons on the video
render_skeleton_video(seq, out_path="skeleton.mp4", blank_canvas=True)      # skeletons on blank canvas

What it does

A functional pipeline over the official Kodokan 100 Techniques YouTube playlist:

acquire (yb) ─► pose (rtmlib / YOLO, tracked tori/uke) ─► segment (motion-energy)
       └─► dol stores (Parquet pose + JSON segments) ─► visualize (overlay / blank / Rerun)
       └─► compare two demos (joint-angle DTW) ─► score + eval harness

YouTube acquisition lives in the yb package (download_youtube_playlist); kodokan is the analysis layer on top.

Install

import kodokan needs only numpy; everything heavy is an optional extra (imported lazily on first use), so the import never fails for a missing one:

pip install -e '.[all]'      # or pick extras: .[pose,viz,analysis,storage,acquire]

You also need ffmpeg on PATH (acquisition/merge). The optional 3D lift (scripts/lift_3d_mediapipe.py) runs in a separate venv, because MediaPipe is ABI-incompatible with numpy 2.x:

python -m venv ~/.kodokan_mp
~/.kodokan_mp/bin/pip install 'mediapipe==0.10.18' 'numpy<2' 'opencv-python-headless==4.10.0.84'

Data (videos, keypoints, renders, weights) lives outside the repo under ~/kodokan_data (override with KODOKAN_DATA_DIR).

The pipeline

Runnable end-to-end examples live in examples/ (warmup_seoinage.py, batch_pipeline.py, segment_review.py, compare_demos.py, score_demos.py, eval_features.py).

Dataset

examples/batch_pipeline.py builds a small dataset (10 techniques · 84 demonstrations · 18.3k frames) into the dol stores. Load it:

from kodokan.store import pose_store, segments_store, load_all_tidy
seq = pose_store()["zIq0xI0ogxk"]          # (F, 2, 17, 3) COCO-17 (x, y, conf)
demos = segments_store()["zIq0xI0ogxk"]    # demo intervals + source_url
df = load_all_tidy()                       # tidy DataFrame across all clips

See misc/docs/dataset.md.

Status & honest limits

Works well: acquisition, tracked two-person pose, demo segmentation, the dol stores, visualization, and same-technique demo comparison (joint-angle DTW is speed-invariant) with interpretable per-joint/per-phase feedback.

Does not work yet — and this is measured, not assumed: technique recognition / cross-demo scoring. A feature bake-off (misc/docs/feature-bakeoff.md) shows every 2D descriptor and MediaPipe 3D joint angles sit at chance (separation AUC ≈ 0.49–0.56). The blockers are noisy monocular 3D under grappling occlusion, tori/uke role inconsistency, and the weakness of hand-crafted angle-DTW for few examples — not viewpoint alone. Recognition needs a learned skeleton representation (few-shot JEANIE, or trained PoseC3D/CTR-GCN) and/or cleaner multi-person 3D with role-consistent features. The eval harness (examples/eval_features*.py) is ready to validate those.

Background & rationale

• misc/docs/research-architecture.md — cited tools/architecture research (75 refs).
• misc/docs/dataset.md — dataset card.
• misc/docs/feature-bakeoff.md — why hand-crafted features don't discriminate techniques (the empirical finding).