File size: 9,650 Bytes
4f3e648 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 | ================================================================================ Huhb3D - Industrial-Grade Synthetic Dataset Commercial Sales Specification Version 1.0 | 2026-05-31 ================================================================================ 1. PRODUCT OVERVIEW -------------------------------------------------------------------------------- Huhb3D is an industrial-grade synthetic dataset designed for training and evaluating 6DoF pose estimation, instance-level segmentation, and CAD topology understanding models. Unlike generic synthetic datasets, Huhb3D provides pixel-perfect ground truth with sub-pixel reprojection error (0.014px), verified by 62 independent quality checks. Three-in-One Ground Truth: - 6DoF Pose Estimation GT (BOP format) - Instance Segmentation GT (COCO + YOLO format) - STEP Topology GT (Feature type per-pixel labeling) 2. INDUSTRIAL-GRADE PRECISION GUARANTEE -------------------------------------------------------------------------------- +---------------------------+------------------+---------------------+ | Metric | Huhb3D | Industry Standard | +---------------------------+------------------+---------------------+ | Reprojection Error | 0.014 px | < 2.0 px | | Camera Intrinsics (fx-fy) | 0.000000 | < 0.01 | | Rotation Matrix Det | 1.0 +/- 0.000001 | |det-1| < 0.01 | | Depth Scale Accuracy | 1.0 (exact) | depth_scale=1.0 | | Depth Precision | 1mm (uint16) | 1-10mm typical | | Near/Far Ratio | 400:1 | 100-1000:1 | | Coordinate Convention | OpenCV standard | OpenCV or OpenGL | | Anti-aliased Masks | No (clean edges) | Often overlooked | +---------------------------+------------------+---------------------+ Key precision features: - Correct OpenGL->OpenCV coordinate transform (Y/Z sign flip) - Depth formula: depth_mm = (near*far)/(far - z_buf*(far-near)) - Rounding (not truncation) for depth quantization - Anti-aliasing disabled during mask rendering for pixel-perfect IDs - Dynamic near/far plane based on camera distance for optimal precision 3. LEGAL COMPLIANCE & COMMERCIAL SAFETY -------------------------------------------------------------------------------- +-------------------------------+-----------------------------------------+ | Risk Category | Huhb3D Solution | +-------------------------------+-----------------------------------------+ | 3D Model Copyright | 100% original, cadquery-generated | | CAD Kernel License | OpenCASCADE LGPL-2.1 (dynamic link OK) | | CAD Library License | cadquery Apache-2.0 (permissive) | | Dataset License | CC-BY-4.0 (commercial use allowed) | | Source Code License | MIT (fully permissive) | | AGPL Contamination | None - fully migrated from AGPL to MIT | | COCO License Field | CC-BY-4.0 (not AGPL) | | Watermark / Restrictions | None | | Third-party Model Usage | Zero - all models procedurally generated | +-------------------------------+-----------------------------------------+ The entire pipeline from model generation to data rendering uses only permissively licensed components. No proprietary or copyleft-contaminated assets are included. The dataset can be used for commercial training, fine-tuning, and redistribution without legal risk. 4. COMPETITIVE ADVANTAGES vs MARKET ALTERNATIVES -------------------------------------------------------------------------------- +---------------------------+------------------+------------------+------------------+ | Feature | Huhb3D | YCB-Video | T-LESS | +---------------------------+------------------+------------------+------------------+ | 6DoF GT Precision | 0.014px reproj | ~1-3px | ~1-2px | | Instance Seg GT | Yes (COCO+YOLO) | No | No | | Feature Type GT | Yes (12 types) | No | No | | Depth Format | uint16 mm | float .mat | uint16 mm | | Coordinate Convention | OpenCV verified | Mixed | OpenCV | | Model Origin | Original CAD | Real objects | Real objects | | STEP Topology | Yes | No | No | | License | CC-BY-4.0 | Research only | Research only | | Commercial Use | Allowed | Restricted | Restricted | | Customizable Pipeline | Yes (open) | No | No | | Anti-aliasing Artifacts | Eliminated | Present | Present | | Depth Scale Consistency | Verified 1.0 | Varies | Varies | +---------------------------+------------------+------------------+------------------+ vs Generic Synthetic Tools (BlenderProc, etc.): - Huhb3D provides verified industrial-grade precision out-of-the-box - No need to manually debug coordinate conventions or depth formulas - Feature-type GT enables topology-aware training (unique in market) - One-click pipeline from model generation to packaged dataset 5. DATASET CONTENTS -------------------------------------------------------------------------------- 20 industrial objects across 8 categories: - Flange (4 variants: default, small, medium, large) - Gear (4 variants: default, small, medium, large) - Bearing (3 variants: default, small, medium) - Bolt, Bracket, Connector, Coupling, Heat Sink - Mounting Plate, Pipe Tee, Shaft, Valve Body Per-object data: - RGB images (800x600, 24-bit PNG) - Depth maps (800x600, 16-bit uint16 PNG, mm units) - Semantic segmentation masks (RGB-encoded feature types) - Instance segmentation masks (RGB=InstanceID,FeatureType,FeatureIdx) - BOP scene_camera.json + scene_gt.json - COCO semantic + instance annotations - YOLO detection + segmentation + instance labels - Feature type legend (12 feature types) 6. COMPATIBILITY -------------------------------------------------------------------------------- +-------------------+-------------------+-----------------------------------+ | Platform | Status | Notes | +-------------------+-------------------+-----------------------------------+ | OpenCV | Fully Compatible | Standard cam_K, R_m2c, t_m2c | | ROS/ROS2 | Fully Compatible | sensor_msgs/Image uint16 depth | | Blender | Compatible | Import via Python script | | PyTorch | Fully Compatible | Standard PNG + JSON loading | | TensorFlow | Fully Compatible | Standard PNG + JSON loading | | BOP Toolkit | Fully Compatible | Standard BOP format | | COCO API | Fully Compatible | Standard COCO with RLE | | YOLOv5/v8/v11 | Fully Compatible | Standard YOLO label format | | Unity Perception | Compatible | Requires format adapter | | Isaac Sim | Compatible | Requires format adapter | +-------------------+-------------------+-----------------------------------+ 7. USE CASES -------------------------------------------------------------------------------- - Industrial robotic bin picking (6DoF pose estimation) - Quality inspection (feature-type segmentation) - Assembly verification (instance-level segmentation) - Sim2Real transfer research (domain randomization ready) - CAD model retrieval (STEP topology matching) - Multi-task learning (6DoF + segmentation + topology) 8. PRICING & LICENSING -------------------------------------------------------------------------------- License: CC-BY-4.0 - Commercial use: Allowed - Redistribution: Allowed with attribution - Modification: Allowed - No warranty: As-is Attribution requirement: "Huhb3D v1.0, Industrial-Grade Synthetic Dataset, CC-BY-4.0" Custom dataset generation services available: - Custom object categories and variants - Higher resolution (up to 4K) - Multi-object scenes with occlusion - Domain randomization (textures, lighting, backgrounds) - Real background compositing for Sim2Real 9. TECHNICAL SPECIFICATIONS -------------------------------------------------------------------------------- Rendering Engine: OpenGL 3.3 Core Profile Depth Buffer: 24-bit (float), converted to uint16 mm Camera Model: Pinhole with per-frame FOV jitter (+/-3 deg) Camera Distance: 800mm (0.8m, industrial inspection range) Near/Far Planes: 40mm / 16000mm (400:1 ratio) Depth Precision: 1mm at object distance Mask Anti-aliasing: Disabled (pixel-perfect instance IDs) Coordinate System: OpenCV (Y-down, Z-forward) Model Unit: Millimeters (mm) ================================================================================ Contact: Huhb3D License: CC-BY-4.0 Version: 1.0.0 ================================================================================ |