DOPE.pytorch

This is an unofficial implementation of DOPE (Deep Object Pose Estimation for Semantic Robotic Grasping of Household Objects) trained on self-created synthetic bottle dataset.

Requirement

$ conda create -n DOPE python=3.6
$ pip install -r requirement.txt
$ conda activate DOPE

Usage

Train and Eval

[WIP]

Demo

• Logs [Download]

 DOPE.pytorch
   - logs
     - Jack_Daniels-checkpoint.pth
     - Jose_Cuervo-checkpoint.pth

• Data

DOPE.pytorch
  - data
    - Real_bottle_sequence
      - 000001.jpg
      - 000002.jpg
      ...
      - _object_settings.json
      - _camera_settings.json

• Run

python demo.py
  --path_to_data_dir ./data/Real_bottle_sequence
  --class_name Jack_Daniels
  --checkpoint ./logs/Jack_Daniels-checkpoint.pth
  --plot

Reference

• DOPE (Deep Object Pose Estimation for Semantic Robotic Grasping of Household Objects) [paper link]
• Some code is borrowed from below repos.
  • Official implementation from nvidia (inference code with ROS) [Deep_Object_Pose]
  • Realtime_Multi-Person_Pose_Estimation [pytorch_Realtime_Multi-Person_Pose_Estimation]