ICP LiDAR ↔ SMPL-X Dense Fusion

Refines Multi-HMR SMPL-X meshes using live iPhone LiDAR via point-to-plane ICP.

Env vars

Var	Default	Effect
`ICP_FUSION`	`0`	`1` enables LiDAR receiver + FusionWorker
`ICP_LIDAR_HOST`	(required when on)	iPhone ARBodyTracker IP on the LAN
`ICP_LIDAR_PORT`	`5500`	TCP port the iOS app publishes ARMesh on
`ICP_LIDAR_EXTRINSIC`	`~/.config/av-live/lidar_extrinsic.json`	Path to persisted extrinsic JSON

Relation to ARKit joint fusion

ICP LiDAR fusion is mesh-level and complementary to the existing joint-level ARKit fusion (iphone_osc_listener.py + pose_filter.py::ArkitFuse + multi_hmr_worker.arkit_pelvis_z_override). The two run independently:

ARKit joints (OSC :57128) — sparse (14 mapped joints), 60 Hz, fast, used to override MediaPipe pose joint slots and lock Multi-HMR pelvis Z.
ICP LiDAR mesh (TCP :5500) — dense (~thousand points), 5–10 Hz, used to register Multi-HMR SMPL-X vertices onto the real-world geometry captured by the iPhone LiDAR.

They can be enabled together or separately. ICP runs only when ICP_FUSION=1.

Calibration

Launch the iPhone ARBodyTracker app and note its LAN IP.

From data_only_viz/:

uv run --extra lidar python -m data_only_viz.scripts.calibrate_lidar \
  --lidar-host <iPhone IP> --lidar-port 5500 --webcam-index 0

The script asks for 4 stances (front / left / right / back). Hold still each time and press ENTER.
The estimated extrinsic is written to ICP_LIDAR_EXTRINSIC (or the default path). Re-run any time the camera or iPhone moves.

NOTE — as of the initial ICP MVP (Task 9), multi_hmr_worker.predict_once is a stub raising NotImplementedError. The calibration CLI runs the LiDAR reader and 4-stance loop scaffold but cannot capture the webcam pelvis side until a follow-up wires predict_once to the existing inference path. Track this in the next planning round.

Runtime

ICP_FUSION=1 ICP_LIDAR_HOST=192.168.0.42 \
  uv run --extra lidar python -m data_only_viz.main

Architecture (summary)

iPhone ARBodyTracker app
  ├── OSC  :57128  /body3d/kp        → IphoneOSCListener (ARKit joint fusion)
  └── TCP  :5500   ARMeshAnchors      → LidarTCPReader  (ICP mesh fusion)
                                                       ↓
                                       FusionWorker.run_once(state)
                                                       ↓
                                       state.persons_smplx[*].vertices_3d
                                       (replaced in place when ICP accepts)

ICP fusion runs in its own daemon thread (IcpFusionThread, target 8 Hz). It is opt-in (off by default) and a no-op if the LiDAR stream is absent.

Troubleshooting

open3d missing → cd data_only_viz && uv sync --extra lidar
No LiDAR frames → check that the iPhone app is publishing on the expected port and that nothing else is bound to it. nc -l 5500 from the Mac should not succeed while the app runs.
ICP always rejected (fitness < 0.30) → the extrinsic is likely stale; re-run calibration. Verify the iPhone is facing the same scene as the webcam.
Mesh appears scaled wrong → SMPL-X is in metres; the iPhone publishes metres. If you see a factor-1000 mismatch the iOS encoder is sending millimetres — patch the iOS app, not this code.
Bench shows latency_ms_p95 > 100 → reduce IcpConfig.voxel_size_m (e.g. 0.03 m) or max_iterations (e.g. 20).
Python cp314 wheel failure on uv sync --extra lidar → open3d does not ship cp313+ wheels yet. Use Python 3.12 (uv venv --python 3.12).

Implementation note (Task 6 deviation)

register_mesh_to_lidar uses a two-stage coarse-to-fine ICP internally: a warm-start pass at max(0.25 m, 5× threshold) correspondence, then a strict pass at IcpConfig.max_correspondence_m (default 0.05 m). The accept/reject gate (fitness ≥ 0.30, rmse ≤ 0.05 m) is evaluated only on the strict pass so the contract is preserved. The warm-start makes ICP converge reliably when Multi-HMR's initial mesh sits more than 5 cm off the LiDAR surface (typical with a fresh calibration or pose change).

4.2 KiB Raw Permalink Blame History Unescape Escape