BlazePose (MediaPipe pose landmarker) — Core ML

Google's MediaPipe Pose Landmarker models (BlazePose GHUM), converted to Core ML for the AML operator family for TouchDesigner. Three variants, same interface:

Interface

• Input image: 256×256 RGB, a square crop around one person (MediaPipe's ROI: centred on the hips, rotated so the body is upright, side 1.25× the body extent). The model does not find people; run a detector first, then track from the previous frame's auxiliary landmarks 33/34 as MediaPipe does.
• Outputs (ML program, fp16; named out0…out4 by the converter — match them by element count):
  • 195 = 39 landmarks × (x, y, z, visibility logit, presence logit), in crop pixels (0..256); landmarks 0..32 are the body, 33..38 auxiliary (33 = ROI centre, 34 = ROI scale point)
  • 1 = pose presence flag
  • 256×256×1 = segmentation mask logits
  • 64×64×39 = heatmaps
  • 117 = 39 world landmarks × (x, y, z) in metres, hips at the origin, y down, z toward the camera

Conversion

pose_landmarker_*.task (float16, 2023-04) → pose_landmarks_detector.tflite → tf2onnx 1.17 (--inputs-as-nchw) → onnx2torch → torch.jit.trace → coremltools 9.0 (mlprogram, FLOAT16, image input scaled 1/255, macOS 13+). Against the TFLite model on the same crop: world landmarks within 0.003 m, 2D landmarks within 1.5 px, pose flag within 0.001. The recipe and the measurements are in the AML repository under labs/blazepose.

Licence

Apache License 2.0 — the same terms as the MediaPipe models these are converted from. Copyright Google LLC (models); conversion by Mickey van Olst, 2026. See LICENSE.