SporeAgent: Reinforced Scene-level Plausibility for Object Pose Refinement
This repository implements the approach described in SporeAgent: Reinforced Scene-level Plausibility for Object Pose Refinement (WACV 2022, arXiv).
Iterative registration using SporeAgent:
The initial pose from PoseCNN (purple) and the final pose using SporeAgent (blue) on the LINEMOD (left,cropped) and
YCB-Video (right) datasets.
Scene-level Plausibility:
The initial scene configuration from PoseCNN (left) results in an implausible pose of the
target object (gray). Refinement using SporeAgent (right) results in a plausible scene configuration where the
intersecting points (red) are resolved and the object rests on its supported points (cyan).
LINEMOD | AD < 0.10d | AD < 0.05d | AD <0.02d | YCB-Video | ADD AUC | AD AUC | ADI AUC |
---|---|---|---|---|---|---|---|
PoseCNN | 62.7 | 26.9 | 3.3 | 51.5 | 61.3 | 75.2 | |
Point-to-Plane ICP | 92.6 | 79.8 | 29.9 | 68.2 | 79.2 | 87.8 | |
w/ VeREFINE | 96.1 | 85.8 | 32.5 | 70.1 | 81.0 | 88.8 | |
Multi-hypothesis ICP | 99.3 | 89.9 | 35.6 | 77.4 | 86.6 | 92.6 | |
SporeAgent | 99.3 | 93.7 | 50.3 | 79.0 | 88.8 | 93.6 |
Comparison on LINEMOD and YCB-Video:
The initial pose and segmentation estimates are computed using PoseCNN. We compare our approach to vanilla Point-to-Plane
ICP (from Open3D), Point-to-Plane ICP augmented by the simulation-based VeREFINE
approach and the ICP-based multi-hypothesis approach used for refinement in PoseCNN.
Dependencies
The code has been tested on Ubuntu 16.04 and 20.04 with Python 3.6 and CUDA 10.2. To set-up the Python environment, use Anaconda and the provided YAML file:
conda env create -f environment.yml --name sporeagent
conda activate sporeagent
.
The BOP Toolkit is additionally required. The BOP_PATH
in
config.py
needs to be changed to the respective clone directory and the packages required by the BOP Toolkit need to
be installed.
The YCB-Video Toolbox is required for experiments on the YCB-Video dataset.
Datasets
We use the dataset versions prepared for the BOP challenge. The required files can be downloaded to a directory of your choice using the following bash script:
export SRC=http://ptak.felk.cvut.cz/6DB/public/bop_datasets
export DATASET=ycbv # either "lm" or "ycbv"
wget $SRC/$DATASET_base.zip # Base archive with dataset info, camera parameters, etc.
wget $SRC/$DATASET_models.zip # 3D object models.
wget $SRC/$DATASET_test_all.zip # All test images.
unzip $DATASET_base.zip # Contains folder DATASET.
unzip $DATASET_models.zip -d $DATASET # Unpacks to DATASET.
unzip $DATASET_test_all.zip -d $DATASET # Unpacks to DATASET.
For training on YCB-Video, the $DATASET_train_real.zip
is moreover required.
In addition, we have prepared point clouds sampled within the ground-truth masks (for training) and the segmentation masks computed using PoseCNN (for evaluation) for the LINEMOD and YCB-Video dataset. The samples for evaluation also include the initial pose estimates from PoseCNN.
LINEMOD
Extract the prepared samples into PATH_TO_BOP_LM/sporeagent/
and set LM_PATH
in config.py
to the base directory,
i.e., PATH_TO_BOP_LM
. Download the PoseCNN results
and the corresponding image set definitions
provided with DeepIM and extract both into POSECNN_LM_RESULTS_PATH
. Finally, since the BOP challenge uses a different
train/test split than the compared methods, the appropriate target file found here
needs to be placed in the PATH_TO_BOP_LM
directory.
To compute the AD scores using the BOP Toolkit, BOP_PATH/scripts/eval_bop19.py
needs to be adapted:
- to use ADI for symmetric objects and ADD otherwise with a 2/5/10% threshold, change
p['errors']
to
{
'n_top': -1,
'type': 'ad',
'correct_th': [[0.02], [0.05], [0.1]]
}
- to use the correct test targets, change
p['targets_filename']
to'test_targets_add.json'
YCB-Video
Extract the prepared samples into PATH_TO_BOP_YCBV/reagent/
and set YCBV_PATH
in config.py
to the base directory,
i.e., PATH_TO_BOP_YCBV
. Clone the YCB Video Toolbox to
POSECNN_YCBV_RESULTS_PATH
. Extract the results_PoseCNN_RSS2018.zip
and copy test_data_list.txt
to the same directory. The POSECNN_YCBV_RESULTS_PATH
in config.py
needs to be changed to the respective directory.
Additionally, place the meshes in the canonical frame models_eval_canonical
in the PATH_TO_BOP_YCBV
directory.
To compute the ADD/AD/ADI AUC scores using the YCB-Video Toolbox, replace the respective files in the toolbox by the
ones provided in sporeagent/ycbv_toolbox
.
Pretrained models
Weights for both datasets can be found here. Download and
copy them to sporeagent/weights/
.
Training
For LINEMOD: python registration/train.py --dataset=lm
For YCB-Video: python registration/train.py --dataset=ycbv
Evaluation
Note that we precompute the normal images used for pose scoring on the first run and store them to disk.
LINEMOD
The results for LINEMOD are computed using the BOP Toolkit. The evaluation script exports the required file by running
python registration/eval.py --dataset=lm
,
which can then be processed via
python BOP_PATH/scripts/eval_bop19.py --result_filenames=PATH_TO_CSV_WITH_RESULTS
.
YCB-Video
The results for YCB-Video are computed using the YCB-Video Toolbox. The evaluation script exports the results in BOP format by running
python registration/eval.py --dataset=ycbv
,
which can then be parsed into the format used by the YCB-Video Toolbox by running
python utility/parse_matlab.py
.
In MATLAB, run evaluate_poses_keyframe.m
to generate the per-sample results and plot_accuracy_keyframe.m
to compute
the statistics.
Citation
If you use this repository in your publications, please cite
@article{bauer2022sporeagent,
title={SporeAgent: Reinforced Scene-level Plausibility for Object Pose Refinement},
author={Bauer, Dominik and Patten, Timothy and Vincze, Markus},
booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year={2022},
pages={654-662}
}