6D Object Pose Estimation using a Particle Filter with Better Initialization

Requirements

CMake >= 3.17
Python >= 3.6
CUDA >= 10.2 (We tested on CUDA 10.2)
Pytorch >= 1.5.0 (We tested on Pytorch 1.5 and 1.7.1 and 1.9)

It currently works on Ubuntu 18.04. You can try this code using docker if your OS is not Ubuntu 18.04.

Installation

Change compile.sh line 5 to the glm library include path. This library can be downloaded from this link.

$ cd lib
$ sh compile.sh
$ cd ..
$ conda env create -f requirements.yaml
$ conda activate pf_with_cpn

Or you can just run below in your own anaconda environment.

$ pip install opencv-python transforms3d open3d scipy

Preparing the datasets and toolbox

YCB Video dataset

Download the YCB Video dataset by following the comments here to your local datasets folder.

Occluded LINEMOD dataset

Download the Occluded LINEMOD dataset from BOP: Benchmark for 6D Object Pose Estimation or you can directly download here to your local datasets folder.

YCB Video toolbox

Download the YCB Video toolbox from here to <local path to 6D_pose_estimation_particle_filter directory>/CenterFindNet/YCB_Video_toolbox directory. And unzip results_PoseCNN_RSS2018.zip.

$ cd <local path to 6D_pose_estimation_particle_filter directory>/CenterFindNet/YCB_Video_toolbox
$ unzip unzip results_PoseCNN_RSS2018.zip

PVNet segmentation results

We generated the segmentation results by using the segmentation network, and pre-trained model from PVNet in off-line. To evaluate the Occluded LINEMOD with PVNet segmentation mask, those segmentation results are required. It can downloaded from here to downloaded Occluded LINEMOD dataset directory. <local path to Occluded LINEMOD dataset>/test/000002/labels_pvnet

Mask R-CNN segmentation results

We generated the segmentation results by using the Mask R-CNN network, and pre-trained model from Pix2Pose in off-line. To evaluate the Occluded LINEMOD with Mask R-CNN segmentation mask, those segmentation results are required. It can downloaded from here to downloaded Occluded LINEMOD dataset directory. <local path to Occluded LINEMOD dataset>/test/000002/labels_mask_rcnn

YCB Video and LIENMOD objects models can be cound in `<local path to 6D_pose_estimation_particle_filter repo>/models`

Demo for 6D pose estimation

You can evaluate the results after saving the estimated pose first.

Runing the demo (for saving the estimated pose results)

Run $ ./save_lmo_estimation.sh for estimating on the Occluded LINEMOD, $ ./save_ycb_estimation.sh for estimating on the YCB Video dataset.

There are 8 options that you can fill out (You must fill out --dataset_root_dir as your datasets local directory.) :

dataset(str) : lmo for Occluded LINEMOD, ycb for YCB Video
dataset_root_dir(str) : <your local path to 6D_pose_estimation_particle_filter directory>/test/000002
save_path(str) : The directory to save the estimated pose. ex) results/lmo/
visualization(bool) : If you don't want to watch how the prediction going on, set this False. Default is True.
gaussian_std(float) : This is Gaussian standard diviation of Eq.(5) in the paper. Default is 0.1.
max_iteration(int) : This value is the maximum number of iterations for an object. Default is 20.
tau(float) : This is the start value of misalignment tolerance τ₀. It is decreased from τ₀ to 0.1*τ₀. Default is 0.1.
num_particles(int) : This is the number of particles. Default is 180.

There is an additional option of choosing the input mask type for the demo of Occluded LINEMOD. We evaluated the 6D pose results on each input segmentation mask of Mask R-CNN and PVNet.

input_mask : Choose between pvnet and mask_rcnn as input mask. Default is pvnet.

Evaluating on the saved results

Run $ ./eval_lmo.sh for estimating on the Occluded LINEMOD, $ ./eval_ycb.sh for estimating on the YCB Video dataset. This step has to be run after the saving pose results. Or you can run with --save_path results/ycb_multi_init_trans/ in eval_ycb.sh line 5 for checking the performance of the our result.

We contain the results of experiments recorded in our paper. Replace the argument of --dataset, --save_path in the file to below.

(Ours - Multi initial translation) :    --dataset ycb --save_path results/ycb_multi_init_trans/
(Ours - 180 particles) :                --dataset ycb --save_path results/ycb_180_particles/
(Ours with Mask R-CNN input) :          --dataset lmo --save_path results/lmo_mask_RCNN_input/
(Ours with PVNet mask input) :          --dataset lmo --save_path results/lmo_PVNet_input/ in our paper.

Demo for Centroid Prediction Network

Training Centroid Prediction Network(CPN)

Run $ ./train_CPN.sh after filling out your local path of the YCB Video dataset in the argument of --dataset_root_dir of the sh file. Weight file would be saved in: <local path to 6D_pose_estimation_particle_filter directory>/CenterFindNet/trained_model/

Evaluating Centroid Prediction Network(CPN)

If you want to evaluate with own your trained model, add the line that argument of --model_path in the eval_CPN_<dataset>.sh. Or you can run directly $ ./eval_CPN_ycb.sh for the YCB Video dataset, $ ./eval_CPN_lmo.sh for the Occluded LINEMOD. If you run it with "--visualization True", the result of projection into the image would be shown on the window.

How to make predefined grasp poses

Convert obj file to pcd file

$ cd lib/obj2pcd && mkdir build && cd build && cmake .. && make
$ ./obj2pcd <obj file path>

Predefine grasp pose

Install graspnetAPI Write your code (grasp pose x,y,z,roll,pitch,yaw ...) in function vis_predefined_grasps

$ python grasp_pose_predefined.py <pcd file path>

wellyowo/Particle_filter_approach_6D_pose_estimation