ngzhili / SynTable

The official code implementation for SynTable - A Synthetic Data Generation Pipeline for Cluttered Tabletop Scenes

SynTable - A Synthetic Data Generation Pipeline for Cluttered Tabletop Scenes

This repository contains the official implementation of the paper "SynTable: A Synthetic Data Generation Pipeline for Unseen Object Amodal Instance Segmentation of Cluttered Tabletop Scenes".

Zhili Ng*, Haozhe Wang*, Zhengshen Zhang*, Francis Eng Hock Tay, Marcelo H. Ang Jr. *equal contributions

[arXiv] [Website] [Dataset] [Demo Video] [Modified UOAIS-v2]

SynTable is a robust custom data generation pipeline that creates photorealistic synthetic datasets of Cluttered Tabletop Scenes. For each scene, it includes metadata such as

• RGB image of scene
• depth image of Scene
• scene instance segmentation masks
• object amodal (visible + invisible) rgb
• object amodal (visible + invisible) masks
• object modal (visible) masks
• object occlusion (invisible) masks
• object occlusion rate
• object visible bounding box
• tabletop visible masks
• background visible mask (background excludes tabletop and objects)
• occlusion ordering adjacency matrix (OOAM) of objects on tabletop

Installation

1. Install NVIDIA Isaac Sim 2022.1.1 version on Omniverse

2. Change Directory to isaac_sim-2022.1.1 directory

   cd '/home/<username>/.local/share/ov/pkg/isaac_sim-2022.1.1/tools'

3. Clone the repo

   git clone https://github.com/ngzhili/SynTable.git

4. Install Dependencies into isaac sim's python

   • Get issac sim source code directory path in command line.

     SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
     echo $SCRIPT_DIR

   • Get isaac sim's python path

     python_exe=${PYTHONEXE:-"${SCRIPT_DIR}/kit/python/bin/python3"}
     echo $python_exe

   • Run isaac sim's python

     $python_exe

   • while running isaac sim's python in bash, install pycocotools and opencv-python into isaac sim's python

     import pip
     package_names=['pycocotools', 'opencv-python'] #packages to install
     pip.main(['install'] + package_names + ['--upgrade'])

5. Copy the mount_dir folder to your home directory (anywhere outside of isaac sim source code)

   cp -r SynTable/mount_dir /home/<username>

Adding object models to nucleus

1. You can download the .USD object models to be used for generating the tabletop datasets here.

2. Upload the downloaded syntable_nucleus folder into Omniverse Nucleus into /Users directory.

3. Ensure that the file paths in the config file are correct before running the generate dataset commands.

Generate Synthetic Dataset

Note: Before generating the synthetic dataset, please ensure that you uploaded all object models to isaac sim nucleus and their paths in the config file is correct.

1. Change Directory to Isaac SIM source code

   cd /home/<username>/.local/share/ov/pkg/isaac_sim-2022.1.1

2. Run Syntable Pipeline (non-headless)

   ./python.sh SynTable/syntable_composer/src/main1.py --input */parameters/train_config_syntable1.yaml --output */dataset/train --mount '/home/<username>/mount_dir' --num_scenes 3 --num_views 3 --overwrite --save_segmentation_data

Types of Flags

Flag	Description
--input	Path to input parameter file.
--mount	Path to mount symbolized in parameter files via '*'.
--headless	Will not launch Isaac SIM window.
--nap	Will nap Isaac SIM after the first scene is generated.
--overwrite	Overwrites dataset in output directory.
--output	Output directory. Overrides 'output_dir' param.
--num-scenes	Number of scenes in dataset. Overrides 'num_scenes' param.
--num-views	Number of views to generate per scene. Overrides 'num_views' param.
--save-segmentation-data	Saves visualisation of annotations into output directory. False by default.

Generated dataset

• SynTable data generation pipeline generates dataset in COCO - Common Objects in Context format.

Folder Structure of Generated Synthetic Dataset

.
├── ...
├── SynTable-Sim                  # Generated dataset
│   ├── data                      # folder to store RGB, Depth, OOAM 
│   │   └── mono
│   │       ├── rgb
│   │       │   ├── 0_0.png       # file naming convention follows sceneNum_viewNum.png
│   │       │   └── 0_1.png 
│   │       ├── depth
│   │       │   ├── 0_0.png  
│   │       │   └── 0_1.png 
│   │       └── occlusion order
│   │           ├── 0_0.npy  
│   │           └── 0_1.npy 
│   ├── parameters                # parameters used for generation of annotations
│   └── train.json                # Annotation COCO.JSON
└── ...

Visualise Annotations

1. Create python venv and install dependencies

   python3.8 -m venv env
   source env/bin/activate
   pip install -r requirements.txt
   

2. Visualise sample annotations (creates a visualise_dataset directory in dataset directory, then saves annotation visualisations there)

   python ./visualize_annotations.py --dataset './sample_data' --ann_json './sample_data/annotation_final.json'
   

Sample Visualisation of Annotations

References

We have heavily modified the Python SDK source code from NVIDA Isaac Sim's Replicator Composer.

Citation

If you find our work useful for your research, please consider citing the following BibTeX entry:

@misc{ng2023syntable,
      title={SynTable: A Synthetic Data Generation Pipeline for Unseen Object Amodal Instance Segmentation of Cluttered Tabletop Scenes}, 
      author={Zhili Ng and Haozhe Wang and Zhengshen Zhang and Francis Tay Eng Hock and Marcelo H. Ang Jr au2},
      year={2023},
      eprint={2307.07333},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}