Bin Picking: Hinge 09
SUMMARY
Physics-simulated, photorealistic RGB renders of metallic hinges in industrial KLT/Euronorm bins, generated with Extreme Domain Randomization (EDR) for sim-to-real transfer, with pixel-accurate instance masks. Access this dataset on Kaggle: Bin Picking Hinge 09
Samples
About the Dataset
A synthetic dataset for instance segmentation of mechanical hinges in cluttered industrial bins, targeting applications in bin-picking, sim-to-real transfer, and amodal/occlusion-heavy segmentation.
Each scene is constructed by spawning 20–30 hinge instances above a randomly selected industrial container and running a rigid-body physics simulation to settle them into a natural, self-occluding pile. The settled scene is then rendered from 4 randomized camera poses under Extreme Domain Randomization (EDR) — aggressive variation of lighting, materials, and backgrounds — yielding photorealistic RGB images with per-instance segmentation masks that transfer well from simulation to real-world inference.
Specifications
| Kaggle link | Bin Picking Hinge 09 |
| Total images | 10,000 |
| Train-Test-Val Split (percentage / number of images) | 70% (6,999) – 10% (1,000) – 20% (2,001) |
| Views per scene | 4 |
| Image resolution | 848 × 480 |
| Image format | PNG (8-bit sRGB) |
| Annotation format | COCO JSON, instance masks as encoded RLE |
| Object classes | 1 (hinge, category_id=34, supercategory=part) |
| Container classes (background) | 1 (bin, category_id=1, supercategory=container) |
| Instances per scene | 20 – 30 (uniform) |
| License | CC BY-NC-SA 4.0 |