trqminh / aistron

Amodal Instance Segmentation Toolbox and Benchmark

Introduction

AIStron is an open-source toolbox that provides current Amodal Instance Segmentation (AIS) methods. It is built as a project using detectron2 (version 0.6) and requires PyTorch 1.8+ or higher. The goal of AIStron is to combine the features of various AIS repositories and align them to facilitate easy maintenance and development of new methods.

Major features:

• Data pipeline: We aim to standardize the annotations of existing AIS datasets so that the methods can be easily implemented and generalized. This diagram illustrates the data pipeline of aistron

• Amodal Instance Segmentation Evaluator: AIStron helps compute the performance of both visible and amodal masks. If a method provides both pred_visible_masks and pred_amodal_masks in its predictions, both performances will be computed. If only the conventional pred_masks are provided, only the amodal performance is computed.

• Utilities: We offer an amodal visualizer utility that allows you to visualize the ground truth or predictions with option to choose between visible masks, occluder masks and amodal masks.

• Builtin-Methods:

  • MaskRCNN

• Methods using aistron as library:

  • BCNet
  • AISFormer
  • ORCNN
  • VRSP-Net

News

• v0.1.0 was released on Jun 19, 2023

Installation

See installation instructions.

Getting Started

• See Getting Started with aistron.
• See Preparing Datasets for aistron.

License

This project is released under the Apache 2.0 license.

Acknowledgements

• We refer to BCNet for dataset mapping with occluder, VRSP-Net for amodal evaluation.
• We base on and detectron2, Mask2Former, and detrex on designing this project.

Citing aistron

If you use aistron in your research, please consider citing this repository using the following BibTeX entry.

@misc{tran2023aistron,
  author =       {Minh Tran},
  title =        {Amodal Instance Segmentation Toolbox and Benchmark},
  howpublished = {\url{https://github.com/trqminh/aistron}},
  year =         {2023}
}