================================================================================ Huhb3D Industrial-Grade Dataset Delivery Checklist Generated: 2026-05-31 13:39:27 Dataset: Huhb3D-Test-Precision-v4 ================================================================================ SECTION 1: MANDATORY FILE STRUCTURE -------------------------------------------------------------------------------- Huhb3D-{version}/ |-- dataset_metadata.json [REQUIRED] Dataset-level metadata |-- LICENSE [REQUIRED] CC-BY-4.0 |-- README.txt [REQUIRED] Usage instructions |-- precision_test_report.txt [REQUIRED] Precision verification report |-- delivery_checklist.txt [REQUIRED] This checklist | |-- {object_name}/ [REQUIRED] One directory per object | |-- scene_camera.json [REQUIRED] BOP camera intrinsics per frame | |-- scene_gt.json [REQUIRED] BOP 6DoF ground truth per frame | |-- label_legend.txt [REQUIRED] Feature type label mapping | | | |-- rgb/ [REQUIRED] RGB images | | |-- rgb_{frame:04d}.png 800x600, 24-bit RGB | | | |-- depth/ [REQUIRED] Depth maps | | |-- depth_{frame:04d}.png 800x600, 16-bit uint16, mm units | | |-- depth_{frame:04d}.npy numpy uint16, mm units | | | |-- mask/ [REQUIRED] Semantic segmentation masks | | |-- mask_{frame:04d}.png 800x600, RGB encoding | | | |-- mask_instance/ [REQUIRED] Instance segmentation masks | | |-- instance_{frame:04d}.png 800x600, RGB=InstanceID,FeatureType,FeatureIdx | | | |-- coco_annotations.json [REQUIRED] COCO semantic format | |-- coco_instance_annotations.json [REQUIRED] COCO instance format | | | |-- yolo_labels/ [REQUIRED] YOLO detection + segmentation | | |-- labels/ | | | |-- {frame:04d}.txt YOLO bbox format | | |-- seg_labels/ | | | |-- {frame:04d}.txt YOLO segmentation format | | | |-- yolo_instance_labels/ [REQUIRED] YOLO instance labels | | |-- {frame:04d}.txt YOLO instance format | | | |-- dataset_report.json [OPTIONAL] Quality assessment report | |-- dataset_report.html [OPTIONAL] HTML quality report | |-- models/ [REQUIRED] 3D model files | |-- {object_name}.stl [REQUIRED] STL mesh for rendering | |-- {object_name}.step [OPTIONAL] STEP CAD model SECTION 2: MANDATORY METADATA - dataset_metadata.json -------------------------------------------------------------------------------- { "dataset_name": "Huhb3D", "version": "1.0.0", "license": "CC-BY-4.0", "description": "6DoF Pose Estimation + Instance Segmentation + STEP Topology GT", "created_date": "YYYY-MM-DD", "model_unit": "mm", "coordinate_convention": "OpenCV (Y-down, Z-forward)", "image_width": 800, "image_height": 600, "fov_degrees": 45.0, "camera_radius_mm": 800.0, "near_plane_mm": 40.0, "far_plane_mm": 16000.0, "depth_format": "uint16 PNG", "depth_unit": "mm", "depth_scale": 1.0, "num_objects": 20, "num_frames_per_object": 10, "total_images": 200, "objects": [ { "name": "flange", "model_file": "models/flange.stl", "step_file": "models/flange.step", "category": "industrial_flange", "bounding_box_mm": [100, 100, 20], "num_vertices": 1234, "num_triangles": 2468 } ], "feature_types": [ {"id": 0, "name": "FreeSurface"}, {"id": 1, "name": "HorizontalPlane"}, {"id": 2, "name": "LateralPlane_X"}, {"id": 3, "name": "LateralPlane_Z"}, {"id": 4, "name": "NearHorizontal"}, {"id": 5, "name": "NearLateral_X"}, {"id": 6, "name": "NearLateral_Z"}, {"id": 7, "name": "Degenerate"}, {"id": 8, "name": "ConvexFeature_Bolt"}, {"id": 9, "name": "ConcaveFeature_Hole"}, {"id": 10, "name": "Flange"}, {"id": 11, "name": "Boss"} ], "generation_pipeline": { "cad_kernel": "OpenCASCADE (LGPL-2.1)", "cad_library": "cadquery 2.x (Apache-2.0)", "rendering_engine": "OpenGL 3.3 Core", "compiler": "MSVC 19.x", "depth_formula": "(near*far)/(far - z_buf*(far-near))", "coordinate_transform": "OpenGL->OpenCV Y/Z sign flip" } } SECTION 3: MANDATORY METADATA - scene_camera.json (per frame) -------------------------------------------------------------------------------- Each frame MUST contain: { "cam_K": [fx, 0, cx, 0, fy, cy, 0, 0, 1], // 3x3 row-major "depth_scale": 1.0, // depth PNG = mm * depth_scale "cam_R_w2c": [...], // 3x3 row-major, world-to-camera rotation "cam_t_w2c": [tx, ty, tz] // 3x1, world-to-camera translation (mm) } Validation rules: - fx MUST equal fy (square pixel) - cx MUST equal imageWidth / 2 - cy MUST equal imageHeight / 2 - depth_scale MUST be 1.0 (depth PNG already stores mm) - cam_R_w2c MUST be valid rotation matrix (det = 1.0) SECTION 4: MANDATORY METADATA - scene_gt.json (per frame) -------------------------------------------------------------------------------- Each frame MUST contain a list of objects: [ { "cam_R_m2c": [...], // 3x3 row-major, model-to-camera rotation "cam_t_m2c": [tx, ty, tz], // 3x1, model-to-camera translation (mm) "obj_id": 1 // Object identifier } ] Validation rules: - cam_R_m2c MUST be valid rotation matrix (det = 1.0) - cam_t_m2c MUST be in mm (same unit as model) - cam_R_m2c = cam_R_w2c * R_model_to_world (for multi-object) - cam_t_m2c = cam_R_w2c * t_model_to_world + cam_t_w2c (for multi-object) SECTION 5: MANDATORY METADATA - Depth Map Specifications -------------------------------------------------------------------------------- Format: 16-bit unsigned integer PNG (uint16) Unit: millimeters (mm) depth_scale: 1.0 (depth_value_mm = pixel_value * depth_scale) Zero value: 0 means no depth (background / beyond far plane) Range: [near_plane_mm, far_plane_mm] Near plane: cameraRadius * 0.05 (mm) Far plane: cameraRadius * 20.0 (mm) Formula: depth_mm = (near * far) / (far - z_buf * (far - near)) z_buf range: [0.0, 1.0] (OpenGL depth buffer value) Rounding: round(depth_mm + 0.5) -> uint16 Precision: sub-pixel reprojection error < 0.1 px SECTION 6: MANDATORY METADATA - Instance Mask Encoding -------------------------------------------------------------------------------- Format: 24-bit RGB PNG Encoding: R = Instance ID (1-255) G = Feature Type ID (0-11) B = Feature Index within type (0-255) Background: RGB = (0, 0, 0) Anti-aliasing: DISABLED during mask rendering (glDisable(GL_MULTISAMPLE)) Feature types: FreeSurface, HorizontalPlane, LateralPlane_X, LateralPlane_Z, NearHorizontal, NearLateral_X, NearLateral_Z, Degenerate, ConvexFeature_Bolt, ConcaveFeature_Hole, Flange, Boss SECTION 7: MANDATORY METADATA - COCO Annotations -------------------------------------------------------------------------------- Semantic (coco_annotations.json): - Standard COCO format with categories from label_legend.txt - RLE encoding: background run first, then alternating fg/bg - Each annotation has: image_id, category_id, segmentation, bbox, area Instance (coco_instance_annotations.json): - Standard COCO format with instance-level annotations - Each annotation has: image_id, category_id, segmentation, bbox, area, iscrowd=0 - category_id maps to object instance (not feature type) SECTION 8: MANDATORY METADATA - YOLO Labels -------------------------------------------------------------------------------- Detection (yolo_labels/labels/): - Format: class_id cx cy w h (normalized 0-1) - One .txt file per image Segmentation (yolo_labels/seg_labels/): - Format: class_id x1 y1 x2 y2 ... (normalized polygon) - One .txt file per image Instance (yolo_instance_labels/): - Format: class_id cx cy w h (normalized, per instance) - One .txt file per image SECTION 9: LEGAL COMPLIANCE CHECKLIST -------------------------------------------------------------------------------- [REQUIRED] License: CC-BY-4.0 (permissive, commercial use allowed) [REQUIRED] 3D models: 100% original, generated by cadquery scripts [REQUIRED] No third-party copyrighted models used [REQUIRED] cadquery: Apache-2.0 license (permissive) [REQUIRED] OpenCASCADE: LGPL-2.1 (dynamic linking OK) [REQUIRED] No AGPL-3.0 dependencies in data pipeline [REQUIRED] COCO license field: CC-BY-4.0 [REQUIRED] No watermarks or restrictions in generated data [REQUIRED] Source code: MIT license SECTION 10: QUALITY ASSURANCE CHECKLIST -------------------------------------------------------------------------------- [PASS] Camera intrinsics: fx == fy (square pixel) [PASS] Principal point: cx = imageWidth/2, cy = imageHeight/2 [PASS] BOP depth_scale: 1.0 [PASS] Depth range: within industrial inspection range (500-2000mm) [PASS] Depth-camera consistency: mean depth ~ camera distance [PASS] Reprojection error: < 0.1 px (sub-pixel) [PASS] Rotation matrix orthogonality: |det(R) - 1| < 0.001 [PASS] Coordinate convention: OpenCV (Y-down, Z-forward) [PASS] Instance masks: anti-aliasing disabled, clean edges [PASS] COCO RLE encoding: standard format (bg_run first) [PASS] YOLO labels: bbox + segmentation + instance [PASS] All 20 objects have complete data (RGB + depth + mask + GT) SECTION 11: DELIVERY VERIFICATION -------------------------------------------------------------------------------- [PASS] bearing_block [PASS] bearing_medium [PASS] bearing_small [PASS] connector_housing [PASS] coupling [PASS] flange [PASS] flange_large [PASS] flange_medium [PASS] flange_small [PASS] gear [PASS] gear_large [PASS] gear_medium [PASS] gear_small [PASS] heat_sink [PASS] hex_bolt [PASS] l_bracket [PASS] mounting_plate [PASS] pipe_tee [PASS] step_shaft [PASS] valve_body Overall: ALL OBJECTS PASS ================================================================================ END OF DELIVERY CHECKLIST ================================================================================