Twenty-one joints, twenty bones

Everything the platform produces — reconstructions, embeddings, features, evidence — is expressed in one coordinate system: a 21-joint, 20-bone skeleton reaching from the pelvis to the ears, with heels and toes included deliberately. Ground contact is where movement quality lives; a topology that stops at the ankle cannot see it.

ear 19·20shoulder 13·16elbow 14·17wrist 15·18hip 3·8heel 6·11nose 2neck 1pelvis 0knee 4·9ankle 5·10toe 7·12
The 21-joint topology, front view. Accent dots are joints; the pelvis (violet) is the root of the kinematic tree. Labels give the canonical joint name and serialization indices (left · right).

Joint ledger

The canonical serialization order, region, and how each joint is sourced from a COCO-17 pose estimate. Row color encodes provenance: direct keypoints, derived midpoints (violet), and extended-estimator joints (amber).

IdxJointRegionCOCO-17 source
0pelviscoremidpoint(left_hip, right_hip)
1neckcoremidpoint(left_shoulder, right_shoulder)
2noseheadnose
3hip_leftleft legleft_hip
4knee_leftleft legleft_knee
5ankle_leftleft legleft_ankle
6heel_leftleft legextended estimator or model inpainting
7toe_leftleft legextended estimator or model inpainting
8hip_rightright legright_hip
9knee_rightright legright_knee
10ankle_rightright legright_ankle
11heel_rightright legextended estimator or model inpainting
12toe_rightright legextended estimator or model inpainting
13shoulder_leftleft armleft_shoulder
14elbow_leftleft armleft_elbow
15wrist_leftleft armleft_wrist
16shoulder_rightright armright_shoulder
17elbow_rightright armright_elbow
18wrist_rightright armright_wrist
19ear_leftheadleft_ear
20ear_rightheadright_ear

Mapping from COCO-17

Most pose estimators emit the 17 COCO keypoints. Fifteen of the twenty-one joints correspond directly; the pelvis and neck are derived as midpoints; heels and toes come from extended estimators that add foot keypoints — or, where a source lacks them, from the model itself, which treats missing joints as one more mask to reconstruct. One consequence is worth stating plainly: when WithoutBall materials mention 17-keypoint pose output, that refers to the upstream computer-vision stage — the product’s own coordinate system is, and remains, these 21 joints.

Why these twenty-one

Each addition to the minimal pose earns its place. Heels and toes expose ground contact, stride events, and push-off mechanics. The pelvis anchors the kinematic tree and separates trunk motion from leg motion. The neck completes the trunk axis, making lean and counter-rotation measurable. Ears stabilize head orientation where broadcast angles hide the face.

The result is the smallest topology that still supports the biomechanical questions the platform exists to answer — asked through the model documented in the architecture reference.