calisthenics-humanoid-decoder-v2

Calisthenics humanoid decoder v2

Compact landmark-to-mesh display decoder for the CaliCombos pose viewer. 283,026 bytes of packed weights; 7,228 vertices and 14,505 triangles. App-owned WASM SIMD execution has a JavaScript fallback. Compatible with the 73-anchor interface of v1.

Refinement

V1 derived the entire head from a shoulder-to-ear neck frame, so nose direction had no effect on the mesh. V2 adds a distinct ear-centre-to-nose head frame, with the ear axis controlling roll. This permits head pitch independently of neck/body posture. Head vertices require a usable face frame; missing or degenerate face anchors are withheld rather than replaced by an invented upward-facing head. The app also stabilizes eyes and ears with the nose.

The surface decoder was distilled from MHR samples and the previously corrected whole-video fit. Body adaptation is blended with the prior decoder and checked against separate synthetic poses to limit regressions. The corrected fit is a training source; its replay is reconstruction, not independent real-person surface validation.

On 64 independent synthetic poses, mean head surface error decreased from 0.02673 to 0.00155 model units. Mean body error decreased from 0.01208 to 0.01199. On corrected-fit reconstruction, mean body error decreased from 0.02277 to 0.01888. These are teacher-relative model-space errors, not measured anatomical accuracy.

A controlled head-pitch test moves the head while leaving all non-head vertices unchanged. V1 had zero head response to that same input. Native/Python, packed JavaScript and WASM paths pass numerical parity checks. The editor is tested in desktop Chrome at desktop and phone-sized layouts; actual phones have not been tested.

This decoder changes mesh display, not raw joints, hold boundaries, video timestamps or skill classifications. It does not locate the face in an image and cannot correct an erroneous input pose. The humanoid is an estimate; individual body shape, clothing silhouette and hidden anatomy are not exact reconstructions.

Runtime

Read manifest.json, verify SHA-256, and unpack the listed tensors. The extra head control is bone 55, [72, 0, 3, 4]: ear centre, nose, left ear, right ear. head_control.requires_face_vertices identifies skull vertices withheld when nose/ears are missing or collinear. All coordinates use x-right, y-up, z-toward. The CaliCombos runtime performs the camera-to-display conversion and registers observed hands to body wrists.

Provenance

MHR surface geometry and SAM 3D Body fit outputs are acknowledged. See LICENSE-MHR.txt and LICENSE-SAM.txt for upstream terms. This is a newly distilled compact decoder, not redistribution of the multi-gigabyte SAM checkpoint. No source videos, review images, per-frame training observations or credentials are included. Validation details are in validation.json.