Seunghyeok Back, Joosoon Lee, Taewon Kim, Sangjun Noh, Raeyoung Kang, Seongho Bak, Kyoobin Lee
This repository contains source codes for the paper "Unseen Object Amodal Instance Segmentation via Hierarchical Occlusion Modeling." (Accepted at ICRA 2022)
[Paper] [Project Website] [Video]
- (2021.09.26) UOAIS-Net has been released
- (2021.11.15) Inference codes for kinect azure and OSD dataset.
- (2021.11.22) ROS nodes for kinect azure and realsense D435
- (2021.12.22) Train and evaluation codes on OSD and OCID dataset + OSD-Amodal annotation
Tested on Titan RTX with python 3.7, pytorch 1.8.0, torchvision 0.9.0, CUDA 10.2 / 11.1 and detectron2 v0.5 / v0.6
- Download source codes and checkpoints
git clone https://github.com/gist-ailab/uoais.git
cd uoais
mkdir output
-
Download checkpoints at GDrive
-
Move the
R50_depth_mlc_occatmask_hom_concat
andR50_rgbdconcat_mlc_occatmask_hom_concat
to theoutput
folder. -
Move the
rgbd_fg.pth
to theforeground_segmentation
folder -
Set up a python environment
conda create -n uoais python=3.8
conda activate uoais
pip install torch torchvision
pip install shapely torchfile opencv-python pyfastnoisesimd rapidfuzz termcolor
-
Install detectron2
-
Build custom AdelaiDet
python setup.py build develop
# UOAIS-Net (RGB-D) + CG-Net (foreground segmentation)
python tools/run_on_OSD.py --use-cgnet --dataset-path ./sample_data --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (depth) + CG-Net (foreground segmentation)
python tools/run_on_OSD.py --use-cgnet --dataset-path ./sample_data --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (RGB-D)
python tools/run_on_OSD.py --dataset-path ./sample_data --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (depth)
python tools/run_on_OSD.py --dataset-path ./sample_data --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
- Realsense D-435 (realsense-ros is required.)
# launch realsense2 driver
roslaunch realsense2_camera rs_aligned_depth.launch
# launch uoais node
roslaunch uoais uoais_rs_d435.launch
# or you can use rosrun
rosrun uoais uoais_node.py _mode:="topic"
- Azure Kinect (Azure_kinect_ROS_Driver is required)
# launch azure kinect driver
roslaunch azure_kinect_ros_driver driver.launch
# launch uoais node
roslaunch uoais uoais_k4a.launch
/uoais/vis_img
(sensor_msgs/Image
): visualization results/uoais/results
(uoais/UOAISResults
): UOAIS-Net predictions (mode:=topic
)/get_uoais_results
(uoais/UOAISRequest
): UOAIS-Net predictions (mode:=service
)
mode
(string
): running mode of ros node (topic
orservice
)rgb
(string
): topic name of the input rgbdepth
(string
): topic name of the input depthcamera_info
(string
): topic name of the input camera infouse_cgnet
(bool
): use CG-Net [1] for foreground segmentation or notuse_planeseg
(bool
): use RANSAC for plane segmentation or notransac_threshold
(float
): max distance a point can be from the plane model
- Realsense D-435 (librealsense and pyrealsense2 are required.)
# UOAIS-Net (RGB-D) + CG-Net (foreground segmentation)
python tools/rs_demo.py --use-cgnet --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
UOAIS-Net (depth) + CG-Net (foreground segmentation)
python tools/rs_demo.py --use-cgnet --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (RGB-D)
python tools/rs_demo.py --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (depth)
python tools/rs_demo.py --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
- Azure Kinect (Azure-Kinect-Sensor-SDK and pyk4a are required.)
# UOAIS-Net (RGB-D) + CG-Net (foreground segmentation)
python tools/k4a_demo.py --use-cgnet --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
UOAIS-Net (depth) + CG-Net (foreground segmentation)
python tools/k4a_demo.py --use-cgnet --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (RGB-D)
python tools/k4a_demo.py --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (depth)
python tools/k4a_demo.py --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
- Download
UOAIS-Sim.zip
andOSD-Amodal-annotations.zip
at GDrive - Download
OSD-0.2-depth.zip
at OSD. [2] - Download
OCID dataset
at OCID. [3] - Extract the downloaded datasets and organize the folders as follows
uoais
├── output
└── datasets
├── OCID-dataset # for evaluation on indoor scenes
│ └──ARID10
│ └──ARID20
│ └──YCB10
├── OSD-0.20-depth # for evaluation on tabletop scenes
│ └──amodal_annotation # OSD-amodal
│ └──annotation
│ └──disparity
│ └──image_color
│ └──occlusion_annotation # OSD-amodal
└── UOAIS-Sim # for training
└──annotations
└──train
└──val
# UOAIS-Net (RGB-D)
python train_net.py --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (depth)
python train_net.py --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (RGB-D) + CG-Net (foreground segmentation)
python eval/eval_on_OSD.py --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml --use-cgnet
# UOAIS-Net (depth) + CG-Net (foreground segmentation)
python eval/eval_on_OSD.py --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml --use-cgnet
This code evaluates the UOAIS-Net that was trained on a single seed (7), thus the metrics from this code and the paper (an average of seeds 7, 77, 777) can be different.
# UOAIS-Net (RGB-D)
python eval/eval_on_OCID.py --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
# UOAIS-Net (depth)
python eval/eval_on_OCID.py --config-file configs/R50_depth_mlc_occatmask_hom_concat.yaml
python tools/run_on_OSD.py --use-cgnet --config-file configs/R50_rgbdconcat_mlc_occatmask_hom_concat.yaml
The source code of this repository is released only for academic use. See the license file for details.
The codes of this repository are built upon the following open sources. Thanks to the authors for sharing the code!
- Instance segmentation based on Detectron2 and AdelaiDet
- Evaluation codes are modified from [4] UCN and [5] VRSP-Net.
- Synthetic dataset is generated with [6] BlenderProc
If you use our work in a research project, please cite our work:
@misc{back2021unseen,
title={Unseen Object Amodal Instance Segmentation via Hierarchical Occlusion Modeling},
author={Seunghyeok Back and Joosoon Lee and Taewon Kim and Sangjun Noh and Raeyoung Kang and Seongho Bak and Kyoobin Lee},
year={2021},
eprint={2109.11103},
archivePrefix={arXiv},
primaryClass={cs.RO}
}
[1] SUN, Yao, et al. Cg-net: Conditional gis-aware network for individual building segmentation in vhr sar images. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-15.
[2] Richtsfeld, Andreas, et al. "Segmentation of unknown objects in indoor environments." 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012.
[3] Suchi, Markus, et al. "EasyLabel: a semi-automatic pixel-wise object annotation tool for creating robotic RGB-D datasets." 2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019.
[4] Xiang, Yu, et al. "Learning rgb-d feature embeddings for unseen object instance segmentation." Conference on Robot Learning (CoRL). 2020.
[5] Xiao, Yuting, et al. "Amodal Segmentation Based on Visible Region Segmentation and Shape Prior." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. No. 4. 2021.
[6] DENNINGER, Maximilian, et al. Blenderproc. arXiv preprint arXiv:1911.01911, 2019.