Zhixuan Zhang, Chi Zhang, Yuehu Liu, Zhenning Niu, Le Wang, Shuqiang Jiang
It is a tool designed for researchers and engineers to annotate scene graph, which is written by python.
- Friendly user interactivity.: It has a friendly GUI. Users can visually view the contents annotated and scene graph in real time.
- Diverse information: In addition to the relationships between instances, it also allows users to annotate diverse information such as regions, clusters and attributes.
- Semi-automatic annotation: GeneAnnotator can recommend the relationship according to the recommendation algorithm and rule system, which greatly improves the annotation efficiency.
- Highly customized: Users are free to define the underlying data, including object categories, relationship lists and so on. It is helpful for researchers to build customize datasets.
- Applying easily: The output dataset files are organized in VG150's data format and can be ready to most existing models for scene graph learning.
Video available on YouTube: https://youtu.be/WeTfNgmp8xA
Video available on Bilibili: https://www.bilibili.com/video/BV1xV411x7cX/
In this annotation software, there are three core modules: Main Annotation Module, Semi-automatic Annotation Module and Auxiliary Function Module. Each module is independent of each other. Removing one of the modules does not affect the use of the whole software, which makes it easy for developers to update and maintain.
25 June 2021: Our paper will be online. https://arxiv.org/abs/2109.02226
25 June 2021: Our software will be launched!
Please refer to INSTALL.md for installation and dataset preparation.
Please see GETTING_STARTED.md for the basic usage.
We support a scene graph dataset with 1000 traffic images by the annotation software. The images are selected from Cityscapes, which is a pixel-level semantic segmented datasets.
In order to show the high variability of traffic scenes, we made efforts to choose images with different features and situations, including crossroad, the tunnel, traffic jam and so on. The steps are: we use places365 to extract the features of the images. Then principal component analysis was used to reduce dimension of features and cluster images into 1000 groups. Images within each group have similar features. Finally, we pick one in each group at random as the original image.
You can download the dataset in https://github.com/Milomilo0320/Traffic-Scene-Graph-1000
This project is released under the Apache 2.0 license.
If you use this toolbox in your research, please cite this paper.
@misc{zhang2021geneannotator,
title={GeneAnnotator: A Semi-automatic Annotation Tool for Visual Scene Graph},
author={Zhixuan Zhang and Chi Zhang and Zhenning Niu and Le Wang and Yuehu Liu},
year={2021},
eprint={2109.02226},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
If you have any questions about our work, please do not hesitate to contact us by emails.
This project is supported by the National Key Research and Development Program of China, No. 2018AAA0102504.