This repository contains the script to generate the HSR-GraspSynth dataset. The dataset is composed of synthetic images of the HSR robot grasping objects in a table. The dataset is generated using BlenderProc and the URDF model of the HSR robot.
To generate the dataset, use the rendering script provided in blenderproc_scripts/. You will need to place the following datasets and files and define the paths in the config file:
- BlenderProc
- URDF model of the HSR robot
- ShapeNet objects
- CC0 Textures
During the training process of the GraspCheckNet model proposed in the paper, the dataset was augmented using the following transformations from the Albumentations library:
| Augmentation | Parameters |
|---|---|
| Perspective Transform | Scale: [0.05, 0.1], p = 0.5 |
| Random Crop and Pad | Percent: [-0.05, 0.1], p = 0.5 |
| Affine Transform | Scale: [1.0, 1.2], p = 0.5 |
| Coarse Dropout | Max Holes: 8, Max Height: 10, Max Width: 10, Min Height: 5, Min Width: 5, p = 0.5 |
| Gaussian Blur | Blur Limit: [3, 7], p = 0.7 |
| Color Jitter | Brightness: [0.8, 1.2], Contrast: [0.2, 1.2], Saturation: [0.2, 0.8], Hue: [-0.1, 0.1], p = 1 |
| Random Brightness-Contrast | Brightness Limit: [-0.25, 0.25], Contrast Limit: [-0.2, 0.2], p = 0.5 |
| Gaussian Noise | Mean: 0, Std Range: [0.05, 0.15], p = 0.5 |
| Optical Distortion | Distort Limit: 0.05, Shift Limit: 0.05, p = 0.5 |
| Image Compression | JPEG Compression |