# ArUco-Based Multi-Camera Extrinsic Calibration from SVO

## TL;DR

> **Quick Summary**: Create a CLI tool that reads synchronized SVO recordings from multiple ZED cameras, detects ArUco markers on a 3D calibration box, computes camera extrinsics using robust pose averaging, and outputs accurate 4x4 transform matrices.
> 
> **Deliverables**:
> - `calibrate_extrinsics.py` - Main CLI tool
> - `pose_averaging.py` - Robust pose estimation utilities
> - `svo_sync.py` - Multi-SVO timestamp synchronization
> - `tests/test_pose_math.py` - Unit tests for pose calculations
> - Output JSON with calibrated extrinsics
> 
> **Estimated Effort**: Medium (3-5 days)
> **Parallel Execution**: YES - 2 waves
> **Critical Path**: Task 1 → Task 3 → Task 5 → Task 7 → Task 8

---

## Context

### Original Request
User wants to integrate ArUco marker detection with SVO recording playback to calibrate multi-camera extrinsics. The idea is to use timestamp-aligned SVO reading to extract frame batches at certain intervals, calculate camera extrinsics by averaging multiple pose estimates, and handle outliers.

### Interview Summary
**Key Discussions**:
- Calibration target: 3D box with 6 diamond board faces (24 markers), defined in `standard_box_markers.parquet`
- Current extrinsics in `inside_network.json` are **inaccurate** and need replacement
- Output: New JSON file with 4x4 pose matrices, marker box as world origin
- Workflow: CLI with preview visualization

**User Decisions**:
- Frame sampling: Fixed interval + quality filter
- Outlier handling: Two-stage (per-frame + RANSAC on pose set)
- Minimum markers: 4+ per frame
- Image stream: Rectified LEFT (no distortion needed)
- Sync tolerance: <33ms (1 frame at 30fps)
- Tests: Add after implementation

### Research Findings
- **Existing patterns**: `find_extrinsic_object.py` (ArUco + solvePnP), `svo_playback.py` (multi-SVO sync)
- **ZED SDK intrinsics**: `cam.get_camera_information().camera_configuration.calibration_parameters.left_cam`
- **Rotation averaging**: `scipy.spatial.transform.Rotation.mean()` for geodesic mean
- **Translation averaging**: Median with MAD-based outlier rejection
- **Transform math**: `T_world_cam = inv(T_cam_marker)` when marker is world origin

### Metis Review
**Identified Gaps** (addressed):
- World frame definition → Use coordinates from `standard_box_markers.parquet`
- Transform convention → Match `inside_network.json` format (T_world_from_cam, space-separated 4x4)
- Image stream → Rectified LEFT view (no distortion)
- Sync tolerance → Moderate (<33ms)
- Parquet validation → Must validate schema early
- Planar degeneracy → Require multi-face visibility or 3D spread check

---

## Work Objectives

### Core Objective
Build a robust CLI tool for multi-camera extrinsic calibration using ArUco markers detected in synchronized SVO playback.

### Concrete Deliverables
- `py_workspace/calibrate_extrinsics.py` - Main entry point
- `py_workspace/aruco/pose_averaging.py` - Robust averaging utilities
- `py_workspace/aruco/svo_sync.py` - Multi-SVO synchronization
- `py_workspace/tests/test_pose_math.py` - Unit tests
- Output: `calibrated_extrinsics.json` with per-camera 4x4 transforms

### Definition of Done
- [ ] `uv run calibrate_extrinsics.py --help` → exits 0, shows required args
- [ ] `uv run calibrate_extrinsics.py --validate-markers` → validates parquet schema
- [ ] `uv run calibrate_extrinsics.py --svos ... --output out.json` → produces valid JSON
- [ ] Output JSON contains 4 cameras with 4x4 matrices in correct format
- [ ] `uv run pytest tests/test_pose_math.py` → all tests pass
- [ ] Preview mode shows detected markers with axes overlay

### Must Have
- Load multiple SVO files with timestamp synchronization
- Detect ArUco markers using cv2.aruco with DICT_4X4_50
- Estimate per-frame poses using cv2.solvePnP
- Two-stage outlier rejection (reprojection error + pose RANSAC)
- Robust pose averaging (geodesic rotation mean + median translation)
- Output 4x4 transforms in `inside_network.json`-compatible format
- CLI with click for argument parsing
- Preview visualization with detected markers and axes

### Must NOT Have (Guardrails)
- NO intrinsic calibration (use ZED SDK pre-calibrated values)
- NO bundle adjustment or SLAM
- NO modification of `inside_network.json` in-place
- NO right camera processing (use left only)
- NO GUI beyond simple preview window
- NO depth-based verification
- NO automatic config file updates

---

## Verification Strategy

> **UNIVERSAL RULE: ZERO HUMAN INTERVENTION**
>
> ALL tasks must be verifiable by agent-executed commands. No "user visually confirms" criteria.

### Test Decision
- **Infrastructure exists**: NO (need to set up pytest)
- **Automated tests**: YES (tests-after)
- **Framework**: pytest

### Agent-Executed QA Scenarios (MANDATORY)

**Verification Tool by Deliverable Type:**

| Type | Tool | How Agent Verifies |
|------|------|-------------------|
| CLI | Bash | Run command, check exit code, parse output |
| JSON output | Bash (jq) | Parse JSON, validate structure and values |
| Preview | Playwright | Capture window screenshot (optional) |
| Unit tests | Bash (pytest) | Run tests, assert all pass |

---

## Execution Strategy

### Parallel Execution Waves

```
Wave 1 (Start Immediately):
├── Task 1: Core pose math utilities
├── Task 2: Parquet loader and validator
└── Task 4: SVO synchronization module

Wave 2 (After Wave 1):
├── Task 3: ArUco detection integration (depends: 1, 2)
├── Task 5: Robust pose aggregation (depends: 1)
└── Task 6: Preview visualization (depends: 3)

Wave 3 (After Wave 2):
├── Task 7: CLI integration (depends: 3, 4, 5, 6)
└── Task 8: Tests and validation (depends: all)

Critical Path: Task 1 → Task 3 → Task 7 → Task 8
```

### Dependency Matrix

| Task | Depends On | Blocks | Can Parallelize With |
|------|------------|--------|---------------------|
| 1 | None | 3, 5 | 2, 4 |
| 2 | None | 3 | 1, 4 |
| 3 | 1, 2 | 6, 7 | 5 |
| 4 | None | 7 | 1, 2 |
| 5 | 1 | 7 | 3, 6 |
| 6 | 3 | 7 | 5 |
| 7 | 3, 4, 5, 6 | 8 | None |
| 8 | 7 | None | None |

---

## TODOs

- [ ] 1. Create pose math utilities module

  **What to do**:
  - Create `py_workspace/aruco/pose_math.py`
  - Implement `rvec_tvec_to_matrix(rvec, tvec) -> np.ndarray` (4x4 homogeneous)
  - Implement `matrix_to_rvec_tvec(T) -> tuple[np.ndarray, np.ndarray]`
  - Implement `invert_transform(T) -> np.ndarray`
  - Implement `compose_transforms(T1, T2) -> np.ndarray`
  - Implement `compute_reprojection_error(obj_pts, img_pts, rvec, tvec, K) -> float`
  - Use numpy for all matrix operations

  **Must NOT do**:
  - Do NOT use scipy in this module (keep it pure numpy for core math)
  - Do NOT implement averaging here (that's Task 5)

  **Recommended Agent Profile**:
  - **Category**: `quick`
    - Reason: Pure math utilities, straightforward implementation
  - **Skills**: []
    - No special skills needed

  **Parallelization**:
  - **Can Run In Parallel**: YES
  - **Parallel Group**: Wave 1 (with Tasks 2, 4)
  - **Blocks**: Tasks 3, 5
  - **Blocked By**: None

  **References**:
  - `py_workspace/aruco/find_extrinsic_object.py:123-145` - solvePnP usage and rvec/tvec handling
  - OpenCV docs: `cv2.Rodrigues()` for rvec↔rotation matrix conversion
  - OpenCV docs: `cv2.projectPoints()` for reprojection

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: rvec/tvec round-trip conversion
    Tool: Bash (python)
    Steps:
      1. python -c "from aruco.pose_math import *; import numpy as np; rvec=np.array([0.1,0.2,0.3]); tvec=np.array([1,2,3]); T=rvec_tvec_to_matrix(rvec,tvec); r2,t2=matrix_to_rvec_tvec(T); assert np.allclose(rvec,r2,atol=1e-6) and np.allclose(tvec,t2,atol=1e-6); print('PASS')"
    Expected Result: Prints "PASS"

  Scenario: Transform inversion identity
    Tool: Bash (python)
    Steps:
      1. python -c "from aruco.pose_math import *; import numpy as np; T=np.eye(4); T[:3,3]=[1,2,3]; T_inv=invert_transform(T); result=compose_transforms(T,T_inv); assert np.allclose(result,np.eye(4),atol=1e-9); print('PASS')"
    Expected Result: Prints "PASS"
  ```

  **Commit**: YES
  - Message: `feat(aruco): add pose math utilities for transform operations`
  - Files: `py_workspace/aruco/pose_math.py`

---

- [ ] 2. Create parquet loader and validator

  **What to do**:
  - Create `py_workspace/aruco/marker_geometry.py`
  - Implement `load_marker_geometry(parquet_path) -> dict[int, np.ndarray]`
    - Returns mapping: marker_id → corner coordinates (4, 3)
  - Implement `validate_marker_geometry(geometry) -> bool`
    - Check all expected marker IDs present
    - Check coordinates are in meters (reasonable range)
    - Check corner ordering is consistent
  - Use awkward-array (already in project) for parquet reading

  **Must NOT do**:
  - Do NOT hardcode marker IDs (read from parquet)
  - Do NOT assume specific number of markers (validate dynamically)

  **Recommended Agent Profile**:
  - **Category**: `quick`
    - Reason: Simple data loading and validation
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: YES
  - **Parallel Group**: Wave 1 (with Tasks 1, 4)
  - **Blocks**: Task 3
  - **Blocked By**: None

  **References**:
  - `py_workspace/aruco/find_extrinsic_object.py:55-66` - Parquet loading with awkward-array
  - `py_workspace/aruco/output/standard_box_markers.parquet` - Actual data file

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: Load marker geometry from parquet
    Tool: Bash (python)
    Preconditions: standard_box_markers.parquet exists
    Steps:
      1. cd /workspaces/zed-playground/py_workspace
      2. python -c "from aruco.marker_geometry import load_marker_geometry; g=load_marker_geometry('aruco/output/standard_box_markers.parquet'); print(f'Loaded {len(g)} markers'); assert len(g) >= 4; print('PASS')"
    Expected Result: Prints marker count and "PASS"

  Scenario: Validate geometry returns True for valid data
    Tool: Bash (python)
    Steps:
      1. python -c "from aruco.marker_geometry import *; g=load_marker_geometry('aruco/output/standard_box_markers.parquet'); assert validate_marker_geometry(g); print('PASS')"
    Expected Result: Prints "PASS"
  ```

  **Commit**: YES
  - Message: `feat(aruco): add marker geometry loader with validation`
  - Files: `py_workspace/aruco/marker_geometry.py`

---

- [ ] 3. Integrate ArUco detection with ZED intrinsics

  **What to do**:
  - Create `py_workspace/aruco/detector.py`
  - Implement `create_detector() -> cv2.aruco.ArucoDetector` using DICT_4X4_50
  - Implement `detect_markers(image, detector) -> tuple[corners, ids]`
  - Implement `get_zed_intrinsics(camera) -> tuple[np.ndarray, np.ndarray]`
    - Extract K matrix (3x3) and distortion from ZED SDK
    - For rectified images, distortion should be zeros
  - Implement `estimate_pose(corners, ids, marker_geometry, K, dist) -> tuple[rvec, tvec, error]`
    - Match detected markers to known 3D points
    - Call solvePnP with SOLVEPNP_SQPNP
    - Compute and return reprojection error
  - Require minimum 4 markers for valid pose

  **Must NOT do**:
  - Do NOT use deprecated `estimatePoseSingleMarkers`
  - Do NOT accept poses with <4 markers

  **Recommended Agent Profile**:
  - **Category**: `unspecified-low`
    - Reason: Integration of existing patterns, moderate complexity
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: NO
  - **Parallel Group**: Wave 2 (after Task 1, 2)
  - **Blocks**: Tasks 6, 7
  - **Blocked By**: Tasks 1, 2

  **References**:
  - `py_workspace/aruco/find_extrinsic_object.py:54-145` - Full ArUco detection and solvePnP pattern
  - `py_workspace/libs/pyzed_pkg/pyzed/sl.pyi:5110-5180` - CameraParameters with fx, fy, cx, cy, disto
  - `py_workspace/svo_playback.py:46` - get_camera_information() usage

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: Detector creation succeeds
    Tool: Bash (python)
    Steps:
      1. python -c "from aruco.detector import create_detector; d=create_detector(); print(type(d)); print('PASS')"
    Expected Result: Prints detector type and "PASS"

  Scenario: Pose estimation with synthetic data
    Tool: Bash (python)
    Steps:
      1. python -c "
         import numpy as np
         from aruco.detector import estimate_pose
         from aruco.marker_geometry import load_marker_geometry
         # Create synthetic test with known geometry
         geom = load_marker_geometry('aruco/output/standard_box_markers.parquet')
         K = np.array([[700,0,960],[0,700,540],[0,0,1]], dtype=np.float64)
         # Test passes if function runs without error
         print('PASS')
         "
    Expected Result: Prints "PASS"
  ```

  **Commit**: YES
  - Message: `feat(aruco): add ArUco detector with ZED intrinsics integration`
  - Files: `py_workspace/aruco/detector.py`

---

- [ ] 4. Create multi-SVO synchronization module

  **What to do**:
  - Create `py_workspace/aruco/svo_sync.py`
  - Implement `SVOReader` class:
    - `__init__(svo_paths: list[str])` - Open all SVOs
    - `get_camera_info(idx) -> CameraInfo` - Serial, resolution, intrinsics
    - `sync_to_latest_start()` - Align all cameras to latest start timestamp
    - `grab_synced(tolerance_ms=33) -> dict[serial, Frame] | None` - Get synced frames
    - `seek_to_frame(frame_num)` - Seek all cameras
    - `close()` - Cleanup
  - Frame should contain: image (numpy), timestamp_ns, serial_number
  - Use pattern from `svo_playback.py` for sync logic

  **Must NOT do**:
  - Do NOT implement complex clock drift correction
  - Do NOT handle streaming (SVO only)

  **Recommended Agent Profile**:
  - **Category**: `unspecified-low`
    - Reason: Adapting existing pattern, moderate complexity
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: YES
  - **Parallel Group**: Wave 1 (with Tasks 1, 2)
  - **Blocks**: Task 7
  - **Blocked By**: None

  **References**:
  - `py_workspace/svo_playback.py:18-102` - Complete multi-SVO sync pattern
  - `py_workspace/libs/pyzed_pkg/pyzed/sl.pyi:10010-10097` - SVO position and frame methods

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: SVOReader opens multiple files
    Tool: Bash (python)
    Preconditions: SVO files exist in py_workspace
    Steps:
      1. python -c "
         from aruco.svo_sync import SVOReader
         import glob
         svos = glob.glob('*.svo2')[:2]
         if len(svos) >= 2:
           reader = SVOReader(svos)
           print(f'Opened {len(svos)} SVOs')
           reader.close()
           print('PASS')
         else:
           print('SKIP: Need 2+ SVOs')
         "
    Expected Result: Prints "PASS" or "SKIP"

  Scenario: Sync aligns timestamps
    Tool: Bash (python)
    Steps:
      1. Test sync_to_latest_start returns without error
    Expected Result: No exception raised
  ```

  **Commit**: YES
  - Message: `feat(aruco): add multi-SVO synchronization reader`
  - Files: `py_workspace/aruco/svo_sync.py`

---

- [ ] 5. Implement robust pose aggregation

  **What to do**:
  - Create `py_workspace/aruco/pose_averaging.py`
  - Implement `PoseAccumulator` class:
    - `add_pose(T: np.ndarray, reproj_error: float, frame_id: int)`
    - `get_inlier_poses(max_reproj_error=2.0) -> list[np.ndarray]`
    - `compute_robust_mean() -> tuple[np.ndarray, dict]`
      - Use scipy.spatial.transform.Rotation.mean() for rotation
      - Use median for translation
      - Return stats dict: {n_total, n_inliers, median_error, std_rotation_deg}
  - Implement `ransac_filter_poses(poses, rot_thresh_deg=5.0, trans_thresh_m=0.05) -> list[int]`
    - Return indices of inlier poses

  **Must NOT do**:
  - Do NOT implement bundle adjustment
  - Do NOT modify poses in-place

  **Recommended Agent Profile**:
  - **Category**: `unspecified-low`
    - Reason: Math-focused but requires scipy understanding
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: YES
  - **Parallel Group**: Wave 2 (with Task 3)
  - **Blocks**: Task 7
  - **Blocked By**: Task 1

  **References**:
  - Librarian findings on `scipy.spatial.transform.Rotation.mean()`
  - Librarian findings on RANSAC-style pose filtering

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: Rotation averaging produces valid result
    Tool: Bash (python)
    Steps:
      1. python -c "
         from aruco.pose_averaging import PoseAccumulator
         import numpy as np
         acc = PoseAccumulator()
         T = np.eye(4)
         acc.add_pose(T, reproj_error=1.0, frame_id=0)
         acc.add_pose(T, reproj_error=1.5, frame_id=1)
         mean_T, stats = acc.compute_robust_mean()
         assert mean_T.shape == (4,4)
         assert stats['n_inliers'] == 2
         print('PASS')
         "
    Expected Result: Prints "PASS"

  Scenario: RANSAC rejects outliers
    Tool: Bash (python)
    Steps:
      1. python -c "
         from aruco.pose_averaging import ransac_filter_poses
         import numpy as np
         # Create 3 similar poses + 1 outlier
         poses = [np.eye(4) for _ in range(3)]
         outlier = np.eye(4); outlier[:3,3] = [10,10,10]  # Far away
         poses.append(outlier)
         inliers = ransac_filter_poses(poses, trans_thresh_m=0.1)
         assert len(inliers) == 3
         assert 3 not in inliers
         print('PASS')
         "
    Expected Result: Prints "PASS"
  ```

  **Commit**: YES
  - Message: `feat(aruco): add robust pose averaging with RANSAC filtering`
  - Files: `py_workspace/aruco/pose_averaging.py`

---

- [ ] 6. Add preview visualization

  **What to do**:
  - Create `py_workspace/aruco/preview.py`
  - Implement `draw_detected_markers(image, corners, ids) -> np.ndarray`
    - Draw marker outlines and IDs
  - Implement `draw_pose_axes(image, rvec, tvec, K, length=0.1) -> np.ndarray`
    - Use cv2.drawFrameAxes
  - Implement `show_preview(images: dict[str, np.ndarray], wait_ms=1) -> int`
    - Show multiple camera views in separate windows
    - Return key pressed

  **Must NOT do**:
  - Do NOT implement complex GUI
  - Do NOT block indefinitely (use waitKey with timeout)

  **Recommended Agent Profile**:
  - **Category**: `quick`
    - Reason: Simple OpenCV visualization
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: YES
  - **Parallel Group**: Wave 2 (with Task 5)
  - **Blocks**: Task 7
  - **Blocked By**: Task 3

  **References**:
  - `py_workspace/aruco/find_extrinsic_object.py:138-145` - drawFrameAxes usage
  - `py_workspace/aruco/find_extrinsic_object.py:84-105` - Marker visualization

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: Draw functions return valid images
    Tool: Bash (python)
    Steps:
      1. python -c "
         from aruco.preview import draw_detected_markers
         import numpy as np
         img = np.zeros((480,640,3), dtype=np.uint8)
         corners = [np.array([[100,100],[200,100],[200,200],[100,200]], dtype=np.float32)]
         ids = np.array([[1]])
         result = draw_detected_markers(img, corners, ids)
         assert result.shape == (480,640,3)
         print('PASS')
         "
    Expected Result: Prints "PASS"
  ```

  **Commit**: YES
  - Message: `feat(aruco): add preview visualization utilities`
  - Files: `py_workspace/aruco/preview.py`

---

- [ ] 7. Create main CLI tool

  **What to do**:
  - Create `py_workspace/calibrate_extrinsics.py`
  - Use click for CLI:
    - `--svo PATH` (multiple) - SVO file paths
    - `--markers PATH` - Marker geometry parquet
    - `--output PATH` - Output JSON path
    - `--sample-interval INT` - Frame interval (default 30)
    - `--max-reproj-error FLOAT` - Threshold (default 2.0)
    - `--preview / --no-preview` - Show visualization
    - `--validate-markers` - Only validate parquet and exit
    - `--self-check` - Run and report quality metrics
  - Main workflow:
    1. Load marker geometry and validate
    2. Open SVOs and sync
    3. Sample frames at interval
    4. For each synced frame set:
       - Detect markers in each camera
       - Estimate pose if ≥4 markers
       - Accumulate poses per camera
    5. Compute robust mean per camera
    6. Output JSON in inside_network.json-compatible format
  - Output JSON format:
    ```json
    {
      "serial": {
        "pose": "r00 r01 r02 tx r10 r11 r12 ty ...",
        "stats": { "n_frames": N, "median_reproj_error": X }
      }
    }
    ```

  **Must NOT do**:
  - Do NOT modify existing config files
  - Do NOT implement auto-update of inside_network.json

  **Recommended Agent Profile**:
  - **Category**: `unspecified-high`
    - Reason: Integration of all components, complex workflow
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: NO
  - **Parallel Group**: Wave 3 (final integration)
  - **Blocks**: Task 8
  - **Blocked By**: Tasks 3, 4, 5, 6

  **References**:
  - `py_workspace/svo_playback.py` - CLI structure with argparse (adapt to click)
  - `py_workspace/aruco/find_extrinsic_object.py` - Main loop pattern
  - `zed_settings/inside_network.json:20` - Output pose format

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: CLI help works
    Tool: Bash
    Steps:
      1. cd /workspaces/zed-playground/py_workspace
      2. uv run calibrate_extrinsics.py --help
    Expected Result: Exit code 0, shows --svo, --markers, --output options

  Scenario: Validate markers only mode
    Tool: Bash
    Steps:
      1. uv run calibrate_extrinsics.py --markers aruco/output/standard_box_markers.parquet --validate-markers
    Expected Result: Exit code 0, prints marker count

  Scenario: Full calibration produces JSON
    Tool: Bash
    Preconditions: SVO files exist
    Steps:
      1. uv run calibrate_extrinsics.py \
           --svo ZED_SN46195029.svo2 \
           --svo ZED_SN44435674.svo2 \
           --markers aruco/output/standard_box_markers.parquet \
           --output /tmp/test_extrinsics.json \
           --no-preview \
           --sample-interval 100
      2. jq 'keys' /tmp/test_extrinsics.json
    Expected Result: Exit code 0, JSON contains camera serials

  Scenario: Self-check reports quality
    Tool: Bash
    Steps:
      1. uv run calibrate_extrinsics.py ... --self-check
    Expected Result: Prints per-camera stats including median reproj error
  ```

  **Commit**: YES
  - Message: `feat(aruco): add calibrate_extrinsics CLI tool`
  - Files: `py_workspace/calibrate_extrinsics.py`

---

- [ ] 8. Add unit tests and final validation

  **What to do**:
  - Create `py_workspace/tests/test_pose_math.py`
  - Test cases:
    - `test_rvec_tvec_roundtrip` - Convert and back
    - `test_transform_inversion` - T @ inv(T) = I
    - `test_transform_composition` - Known compositions
    - `test_reprojection_error_zero` - Perfect projection = 0 error
  - Create `py_workspace/tests/test_pose_averaging.py`
  - Test cases:
    - `test_mean_of_identical_poses` - Returns same pose
    - `test_outlier_rejection` - Outliers removed
  - Add `scipy` to pyproject.toml if not present
  - Run full test suite

  **Must NOT do**:
  - Do NOT require real SVO files for unit tests (use synthetic data)

  **Recommended Agent Profile**:
  - **Category**: `quick`
    - Reason: Straightforward test implementation
  - **Skills**: []

  **Parallelization**:
  - **Can Run In Parallel**: NO
  - **Parallel Group**: Wave 3 (final)
  - **Blocks**: None
  - **Blocked By**: Task 7

  **References**:
  - Task 1 acceptance criteria for test patterns
  - Task 5 acceptance criteria for averaging tests

  **Acceptance Criteria**:

  **Agent-Executed QA Scenarios:**

  ```
  Scenario: All unit tests pass
    Tool: Bash
    Steps:
      1. cd /workspaces/zed-playground/py_workspace
      2. uv run pytest tests/ -v
    Expected Result: Exit code 0, all tests pass

  Scenario: Coverage check
    Tool: Bash
    Steps:
      1. uv run pytest tests/ --tb=short
    Expected Result: Shows test results summary
  ```

  **Commit**: YES
  - Message: `test(aruco): add unit tests for pose math and averaging`
  - Files: `py_workspace/tests/test_pose_math.py`, `py_workspace/tests/test_pose_averaging.py`

---

## Commit Strategy

| After Task | Message | Files | Verification |
|------------|---------|-------|--------------|
| 1 | `feat(aruco): add pose math utilities` | pose_math.py | python import test |
| 2 | `feat(aruco): add marker geometry loader` | marker_geometry.py | python import test |
| 3 | `feat(aruco): add ArUco detector` | detector.py | python import test |
| 4 | `feat(aruco): add multi-SVO sync` | svo_sync.py | python import test |
| 5 | `feat(aruco): add pose averaging` | pose_averaging.py | python import test |
| 6 | `feat(aruco): add preview utils` | preview.py | python import test |
| 7 | `feat(aruco): add calibrate CLI` | calibrate_extrinsics.py | --help works |
| 8 | `test(aruco): add unit tests` | tests/*.py | pytest passes |

---

## Success Criteria

### Verification Commands
```bash
# CLI works
uv run calibrate_extrinsics.py --help  # Expected: exit 0

# Marker validation
uv run calibrate_extrinsics.py --markers aruco/output/standard_box_markers.parquet --validate-markers  # Expected: exit 0

# Tests pass
uv run pytest tests/ -v  # Expected: all pass

# Full calibration (with real SVOs)
uv run calibrate_extrinsics.py --svo *.svo2 --markers aruco/output/standard_box_markers.parquet --output calibrated.json --no-preview
jq 'keys' calibrated.json  # Expected: camera serials
```

### Final Checklist
- [ ] All "Must Have" present
- [ ] All "Must NOT Have" absent
- [ ] All tests pass
- [ ] CLI --help shows all options
- [ ] Output JSON matches inside_network.json pose format
- [ ] Preview shows detected markers with axes