akeaveny/densefusion

6-DoF Object Based Pose Estimation in Pytorch

DenseFusion

This work forks DenseFusion, to work with a custom dataset I created. Here is the link to the original paper for DenseFusion.

I used densefusion with the following repos:

1. LabelFusion for generating real images.
2. NDDS for generating synthetic images.
3. arl-affpose-dataset-utils a custom dataset that I generated.
4. pytorch-simple-affnet for predicting an object affordance labels.
5. arl-affpose-ros-node: for deploying our network for 6-DoF pose estimation with our ZED camera.
6. barrett-wam-arm for robotic grasping experiments. Specifically barrett_tf_publisher and barrett_trac_ik.

In the sample below we see real time implementation on our 7-DoF Robot Arm.

Requirements

conda env create -f environment.yml --name DenseFusion

AffNet

1. To inspect ground truth object pose (first look at relative paths for root folder of dataset in tools/ARLAffPose/cfg.py):

   python tools/ARLAffPose/scripts/load_gt_obj_poses.py
   

2. To inspect ground truth affordance pose (see PyTorch-Simple-AffNet):

   python tools/ARLAffPose/scripts/load_gt_obj_part_poses.py
   

3. To run training:

   python tools/train.py
   

4. To get predicted pose run:

   python tools/ARLAffPose/scripts/evaluate_poses_keyframe.py