Active Pointly-Supervised Instance Segmentation (APIS)

Code for the paper "Active Pointly-Supervised Instance Segmentation", ECCV 2022.

NOTE: This release is currently a preliminary version for APIS, where only the newly added or modified source files are included for simplicity. The provided scripts could help you understand how APIS works. We will release the complete version as well as the checkpoints in the near future.

Preparation

This project is based on the open-source toolbox AdelaiDet (as well as Detectron2).

Please refer to INSTALL.md for installation and dataset (MS-COCO) preparation.

The expected folder structure:

ROOT_PATH
├── AdelaiDet
│   ├── datasets
│   │   ├── coco
│   │   │   ├── annotations
│   │   │   ├── train2017
│   │   │   ├── val2017
├── detectron2
├── APIS
│   ├── scripts
│   ├── src

Note that only the newly added or modified source files are included in APIS/src.

Set $ROOT_PATH in APIS/scripts/prepare.sh and run:

# copy source files and prepare random point annotations
sh APIS/scripts/prepare.sh

Usage

We provide the one-click scripts to reproduce the main results in the paper, including the results of the Random Sampling and Entropy strategies mentioned in the paper.

1. model initialization (P0)

python APIS/scripts/initialization.py

2. random sampling (P1~P9)

python APIS/scripts/random.py

3. active selection with the Entropy metric (P1~P9)

python APIS/scripts/entropy.py

Reference

If this work is useful to your research, please cite:

@inproceedings{tang2022APIS,
  title={Active Pointly-Supervised Instance Segmentation},
  author={Tang, Chufeng and Xie, Lingxi and Zhang, Gang and Zhang, Xiaopeng and Tian, Qi and Hu, Xiaolin},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}