SyMFM6D: Symmetry-aware Multi-directional Fusion for Multi-View 6D Object Pose Estimation

This is the official source code for the paper "SyMFM6D: Symmetry-aware Multi-directional Fusion for Multi-View 6D Object Pose Estimation" by Fabian Duffhauss et al. accepted to the IEEE Robotics and Automation Letters (RA-L) 2023. The code allows the users to reproduce and extend the results reported in the study. Please cite the paper when reporting, reproducing, or extending the results.

Installation

• Install CUDA 10.1

• Create a virtual environment with all required packages:

  conda create -n SyMFM6D python=3.6
  conda activate SyMFM6D
  conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.1 -c pytorch
  conda install -c anaconda -c conda-forge scikit-learn
  conda install pytorch-lightning==1.4.9 torchmetrics==0.6.0 -c conda-forge
  conda install matplotlib einops tensorboardx pandas opencv==3.4.2 -c conda-forge
  pip install opencv-contrib-python==3.4.2.16

• Compile the RandLA-Net operators:

  cd models/RandLA/
  sh compile_op.sh

• Download and install normalSpeed:

  git clone https://github.com/hfutcgncas/normalSpeed.git
  cd normalSpeed/normalSpeed
  python3 setup.py install

Datasets and Models

• The YCB-Video dataset can be downloaded here.
• The MV-YCB SymMovCam dataset can be downloaded here. Using this dataset requires the 3D models of the YCB-Video dataset which can be downloaded here.

After downloading a new dataset, the zip file needs to be extracted and the paths to the 3D models and the datasets need to be modified in common.py. For training a new model from scratch, a ResNet-34 pre-trained on ImageNet is required which can be downloaded here.

Finding Object Symmetries

Our symmetry-aware training procedure requires the rotational symmetry axes of all objects. We compute them once in advance of the training by running the script find_symmetries.py for each object. For objects with multiple symmetry axes, the script needs to be ran multiple times with different initial values for the symmetry axis, e.g.

python utils/find_symmetries.py --obj_name 024_bowl --symmtype rotational

We also provide pre-computed symmetry axes for all objects which we used for the paper in symmetries.txt.

Training Models

In the following, we give a few examples how to train models with our SyMFM6D approach on the different datasets.

Single-View Training on YCB-Video

python run.py --dataset ycb --workers 4 --run_name YcbVideo_1view_training --epochs 40 --batch_size 9 \
    --sift_fps_kps 1 --symmetry 1 --n_rot_sym 16

Multi-View Training on YCB-Video

python run.py --dataset ycb --workers 4 --run_name YcbVideo_3views_training --epochs 10 --batch_size 3 \
    --sift_fps_kps 1 --symmetry 1 --n_rot_sym 16 --multi_view 1 --set_views 3  --checkpoint single_view_checkpoint.ckpt \
    --lr_scheduler reduce --lr 7e-05

Single-View Training on MV-YCB SymMovCam

python run.py --dataset SymMovCam --workers 4 --run_name MvYcbSymMovCam_1view_training --epochs 60 \
    --batch_size 3 --sift_fps_kps 1 --symmetry 1 --n_rot_sym 16

Multi-View Training on MV-YCB SymMovCam

python run.py --dataset SymMovCam --workers 4 --run_name MvYcbSymMovCam_1view_training --epochs 60 \
    --batch_size 3 --sift_fps_kps 1 --symmetry 1 --n_rot_sym 16 --multi_view 1 --set_views 3

Evaluating Models

A model can be evaluated by specifying a checkpoint using --checkpoint <name_of_checkpoint> and by adding the argument --test, e.g.

python run.py --dataset ycb --workers 4 --run_name YcbVideo_3views_evaluation --batch_size 3 --sift_fps_kps 1 \
    --multi_view 1 --set_views 3 --checkpoint YcbVideo_3views_checkpoint.ckpt 

Runtime

We provide the runtime of our SyMFM6D approach for different number of views in the following table:

Number of Views	Network Forward Time	Pose Estimation Time	Total Time
1	46 ms	14 ms	60 ms
2	92 ms	19 ms	111 ms
3	138 ms	25 ms	163 ms
4	184 ms	30 ms	214 ms
5	230 ms	36 ms	266 ms

The network forward time includes the 3D keypoint offset prediction, the center point offset prediction, and the prediction of semantic labels. The pose estimation time represents the time for applying the mean shift clustering algorithm and the least-squares fitting for computing the 6D pose of a single object. Please note that the usage of the symmetry-aware loss does increase the training time slightly, but it does not affect the runtime. The runtimes are measured using a single GPU of type NVIDIA Tesla V100 with 32 GB of memory.

Network Parameters

In the following, we show the tensor shapes of our network architecture:

Layer	CNN Tensor Shape	PCN Tensor Shape
1	H/4, W/4, 64	N, 8
2	H/4, W/4, 64	N/4, 64
3	H/8, W/8, 128	N/16, 128
4	H/8, W/8, 512	N/64, 256
5	H/8, W/8, 1024	N/256, 512
6	H/4, W/4, 256	N/64, 256
7	H/2, W/2, 64	N/16, 128
8	H/2, W/2, 64	N/4, 64

The multi-layer perceptrons (MLP) of our multi-directional fusion modules have the following channel sizes, where c_i and c_p denote the channel sizes of the image and point features of the respective layer:

MLP	Channel Sizes
MLPi	ci, cp
MLPfi	2cp, cp
MLPp	cp, ci
MLPfp	2ci, ci

The channel sizes of the MLPs in the second stage of our network in Fig. 2 of our paper are as followed:

MLP	Channel Sizes
Keypoint Detection	ci + cp, 128, 128, 128, ncls
Center Point Detection	ci + cp, 128, 128, 128, 3nkps
Semantic Segmentation	ci + cp, 128, 128, 128, 3

For the output of the network architecture we set c_i = c_p = 64. n_cls is the number of classes in the used dataset, e.g. n_cls = 22 for YCB-Video and SymMovCam. n_kps is the number of keypoints per object, which we set to eight.

License

The code of SyMFM6D is open-sourced under the AGPL-3.0 license. See the LICENSE file for details. The MV-YCB SymMovCam dataset is open-sourced under the CC-BY-SA-4.0 license.

For a list of other open source components included in SyMFM6D, see the file 3rd-party-licenses.txt.

Citation

If you use this work please cite:

@article{Duffhauss_2023,
    author    = {Duffhauss, Fabian and Koch, Sebastian and Ziesche, Hanna and Vien, Ngo Anh and Neumann, Gerhard},
    title     = {SyMFM6D: Symmetry-aware Multi-directional Fusion for Multi-View 6D Object Pose Estimation},
    year      = {2023},
    journal   = {IEEE Robotics and Automation Letters (RA-L)}
}