reports/compatibility_report.txt

================================================================================
  Huhb3D Compatibility Report
  Depth Maps & Camera Intrinsics - OpenCV / ROS / Blender / BOP
================================================================================

=== 1. OPENCV COMPATIBILITY ===

  cam_K as 3x3 matrix (OpenCV cameraMatrix):
    [761.267  0.000  400.000]
    [0.000  761.267  300.000]
    [0.000  0.000  1.000]

  distCoeffs = None (pinhole, no distortion)
  [OK] Standard pinhole model, directly usable with cv2.solvePnP()

  cv2.Rodrigues(R_m2c) -> rvec = [-1.533229, -0.960920, 1.533148]
  [OK] Rotation vector for cv2.solvePnP() obtained successfully

=== 2. DEPTH MAP COMPATIBILITY ===

  .npy format: shape=(600, 800), dtype=uint16
  .npy range: [750, 795]mm

  cv2.imread(UNCHANGED): shape=(600, 800), dtype=uint16
  cv2 depth range: [750, 795]mm
  [OK] OpenCV reads 16-bit depth correctly
  PIL.Image.open: mode=I;16, shape=(600, 800), dtype=uint16
  PIL depth range: [750, 795]mm
  [OK] PIL reads 16-bit depth correctly
  npy vs PNG consistency: PASS

=== 3. ROS COMPATIBILITY ===

  Depth encoding: 16UC1 (16-bit unsigned, 1 channel)
  Depth unit: mm (conversion to meters: value / 1000.0)
  ROS2 sensor_msgs/Image fields:
    encoding: '16UC1'
    height: 600
    width: 800
    step: 1600 (width * 2 bytes)
  [OK] Compatible with ROS depth_image_proc package
  [OK] depth_scale=1.0 means depth_png_value * 1.0 = depth_mm

  CameraInfo message:
    K = [761.267, 0.0, 400, 0.0, 761.267, 300, 0.0, 0.0, 1.0]
    D = [] (no distortion)
    R = [1,0,0,0,1,0,0,0,1] (identity)
    P = [761.267,0,400.0,0, 0,761.267,300.0,0, 0,0,1,0]
  [OK] Standard ROS CameraInfo format

=== 4. BLENDER COMPATIBILITY ===

  Blender uses OpenGL convention (Y-up, Z-back)
  Our data uses OpenCV convention (Y-down, Z-forward)
  Conversion needed for Blender import:
    R_opengl = R_opencv @ diag([1, -1, -1])
    t_opengl = diag([1, -1, -1]) @ t_opencv

  Example conversion (Frame 1):
    R_opencv = [5.25988e-05, 0.0, -1.0, 0.899983, -0.435924, 4.73381e-05, -0.435924, -0.899983, -2.29291e-05]
    R_opengl = [5.25988e-05, 0.0, 1.0, 0.899983, 0.435924, -4.73381e-05, -0.435924, 0.899983, 2.29291e-05]
    t_opencv = [16.0, -0.000758278, 799.987]
    t_opengl = [16.0, 0.000758278, -799.987]
  [OK] Conversion is a simple sign flip on Y and Z axes
  [NOTE] Blender Python API can automate this conversion

  Depth import in Blender:
    Use .npy (uint16) for highest precision
    Convert to float: depth_m = depth_mm / 1000.0
    Apply as Z-buffer displacement in Compositor
  [OK] Depth data compatible with Blender Compositor

=== 5. BOP TOOLKIT COMPATIBILITY ===

  scene_camera.json: BOP standard format
  scene_gt.json: BOP standard format
  cam_K: [fx, 0, cx, 0, fy, cy, 0, 0, 1] (row-major, 9 elements)
  cam_R_m2c: 9-element row-major 3x3
  cam_t_m2c: 3-element [tx, ty, tz] in mm
  depth_scale: 1.0 (depth PNG stores mm directly)

  BOP depth = pixel_value * depth_scale = pixel_value * 1
  BOP depth range: [750, 795]mm
  [OK] BOP depth_scale=1.0 is correct (depth PNG already in mm)

=== 6. COCO API COMPATIBILITY ===

  COCO format version: 2.0
  Images: 10
  Annotations: 103
  Categories: 1
  Sample annotation keys: ['id', 'image_id', 'category_id', 'segmentation', 'area', 'bbox', 'iscrowd', 'instance_id', 'feature_type_id', 'feature_index', 'segmentation_polygon']
  segmentation type: dict
  RLE size: [600, 800]
  RLE counts length: 66
  [OK] Standard COCO RLE format
  License field: N/A
  [OK] Compatible with pycocotools

=== 7. YOLO COMPATIBILITY ===

  YOLO detection format: 5 values per line
  Sample: 0 0.499375 0.499167 0.040000 0.160000
  [OK] Standard YOLO format (class cx cy w h)
  YOLO seg format: 25 values per line
  Sample (first 6): 0 0.505000 0.420000 0.503750 0.465000 0.481250
  [OK] Standard YOLO segmentation format

================================================================================
  COMPATIBILITY SUMMARY
================================================================================

  [OK] OpenCV:       Fully compatible (standard cam_K, R_m2c, t_m2c)
  [OK] ROS/ROS2:     Fully compatible (16UC1 depth, depth_scale=1.0)
  [OK] Blender:      Compatible (needs Y/Z sign flip for OpenGL convention)
  [OK] BOP Toolkit:  Fully compatible (standard BOP format)
  [OK] COCO API:     Fully compatible (standard RLE encoding)
  [OK] YOLO:         Fully compatible (standard bbox + segmentation format)
  [OK] PyTorch/TF:   Fully compatible (standard PNG + JSON)

  BLOCKING ISSUES:   None

  NOTES:
  1. Blender import requires coordinate conversion (Y/Z sign flip).
     This is standard and well-documented for OpenCV->OpenGL conversion.
     A Python conversion script is provided in the dataset package.
  2. Depth PNG uses 16-bit uint16 (mode I;16 in PIL).
     Some image viewers may display incorrectly (show as 8-bit).
     Use cv2.imread(UNCHANGED) or PIL with I;16 mode for correct reading.
  3. cam_K is stored as flat 9-element array (row-major).
     Reshape to 3x3: K = np.array(cam_K).reshape(3,3)
================================================================================