Utilities for working with Contact-GraspNet, a deep learning model for generating 6-DoF grasps from RGB-D input.
These utilities are also helpful for getting started with creating segmentation masks with the Unseen Object Instance Segmentation for Robotic Enviroments framework. These segmentation masks improve the quality of grasps generated from Contact-GraspNet.
Both Contact-Graspnet and UOIS require additional setup which is specified in their own respective repositories.
This repo provides helper tools for:
- Collecting RGB and RGB-D data from an Intel RealSense Depth Camera D435
- Converting RGB and RGB-D data into .npy format for Contact-GraspNet (and UOIS)
- Installation scripts for Contact-Graspnet (in progress)
-
Installation requirements (in progress)
-
If you do not already have a camera intrinsics matrix for your current camera setup. Generate by running:
python image_capture_intel_realsense/camera_intrinsics.py mv image_capture_intel_realsense/camera_matrix.txt input_conversion_tools/input_data
-
Capture RGB and RGB-D data:
python image_capture_intel_realsense/capture_depth_and_rgb_images.py mv image_capture_intel_realsense/scenes input_conversion_tools/input_data
-
Convert scenes into .npy files. This is the file type expected from Contact-GraspNet and UOIS
python input_conversion_tools/image_and_matrix_to_npy.py
Files in processed data can the be use in Contact-GraspNet and UOIS