SBPL-Cruz / perception

PERCH 2.0 : Fast and Accurate GPU-based Perception via Search for Object Pose Estimation
https://sbpl-cruz.github.io/perception/
BSD 3-Clause "New" or "Revised" License
16 stars 5 forks source link

PERCH 2.0 : Fast and Accurate GPU-based Perception via Search for Object Pose Estimation

Image of 6-Dof

Overview

This library provides implementations for single and multi-object 3-Dof and 6-Dof pose estimation from RGB-D sensor data and 3D CAD models. It can evaluate thousands of poses in parallel on a GPU in order to find the pose that best explains the input scene using CUDA. Each pose is refined in parallel through CUDA based GICP. PERCH 2.0 works in conjunction with an instance segmentation CNN for 6-Dof pose estimation (Tested with YCB Video Dataset).

The libray is the official implementation of "PERCH 2.0 : Fast and Accurate GPU-based Perception via Search for Object Pose Estimation" accepted at IROS 2020 [PDF]

Notable Features

System Requirements

Docker Setup

Follow the steps outlined in this Wiki to setup the code on your machine. The code will be built and run from the Docker image and no local installation of dependencies (apart from NVidia drivers and Docker itself) is needed.

Running with YCB Video Dataset

Follow the steps outlined in this Wiki to run the code on YCB Video Dataset. It can run using PoseCNN masks, ground truth masks or a custom MaskRCNN model trained by us. The model is trained to detect full bounding boxes and instance segmentation masks of YCB objects in the dataset.

Results :

Running with Robot

PERCH 2.0 communicates with the robot's camera using ROS. Follow the steps outlined in this Wiki to first test the code with bagfiles. You can then use the bagfile setup of your choice and modify it as per the robot requirements.

Author

Created by Aditya Agarwal at the Search Based Planning Lab, Robotics Institute, Carnegie Mellon University. Please direct any questions about the code or paper to the Issues section of this repo.

Citation

Please use the citation below for our paper accepted at IROS 2020, if you use our code :

@inproceedings{Agarwal2020PERCH2,
  title={PERCH 2.0 : Fast and Accurate GPU-based Perception via Search for Object Pose Estimation},
  author={Aditya Agarwal and Yupeng Han and Maxim Likhachev},
  year={2020},
  booktitle={IROS}
}