Skip to content

Latest commit

 

History

History
71 lines (59 loc) · 3.55 KB

README.md

File metadata and controls

71 lines (59 loc) · 3.55 KB
Raw

Learning Orientation Distributions for Object Pose Estimation

Created by Brian Okorn at R-PaD lab at the Carnegie Mellon Robotics Institute.

Overview

We introduce two learned methods for estimating a distribution over an object's orientation. Our methods take into account both the inaccuracies in the pose estimation as well as the object symmetries. Our first method, which regresses from deep learned features to an isotropic Bingham distribution, gives the best performance for orientation distribution estimation for non-symmetric objects. Our second method learns to compare deep features and generates a non-parameteric histogram distribution. This method gives the best performance on objects with unknown symmetries, accurately modeling both symmetric and non-symmetric objects, without any requirement of symmetry annotation. Project, arXiv.

Citation

If you find our work useful, please consider citing:

@inproceedings{okorn2020learning,
    Author = {Okorn, Brian and Xu, Mengyun and Hebert, Martial and Held, David },
    Title = {Learning Orientation Distributions for Object Pose Estimation},
    Journal = {International Conference on Intelligent Robots and Systems (IROS)},
    Year = {2020}
}

Requirements

  • pytorch 1.7.0
  • numpy
  • PIL
  • scipy
  • tqdm

quat_math

Library of orientation helper functions as well as a wrapper to Christoph Gohlke's original transform library. To install, pip install in the root directory

pip install .

object_pose_utils

Utility functions, datasets, and wrapper to Julian Straub's 4D spherical descritization code To install, pip install in the root directory

pip install .

pybingham

To use the Bingham distribution code, install our our python wrapper to the Bingham Statistics Library from our fork. Follow the python install instructions here.

DenseFusion

To install Dense Fusion as a stand alone library, generate new features, and to have access to the dropout version described in the paper, install our fork. To install, pip install in the root directory

pip install .

PoseCNN

To generate features for PoseCNN, use our fork and this tool or the notebook.

Installation

This code should be installed as a stand alone library using pip in the root directory.

pip install .

Datasets

Download the YCB-Video dataset from here

Pretrained Weights and Feature Grid

Download our pretrained models and generated feature grids here.

Usage

See notebooks/ for interactive examples of using our models and datasets.

Training

See training_scripts/ for example scripts for training orientation distribution networks.

Renderer

To render models for grids, you can use our stand alone Blender renderer or pyrender.