biodatlab / icip-challenge-2022

Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022 (3rd place solution)

Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022

This repository hosts scripts for ICIP 2022 challenge for parasitic egg detection and classification in Microscopic images solution developed by Biomedical and Data Lab at Mahidol University, Thailand. You can see the challenge website here.

Our technique applies:

• mulitask learning using pseudo mask generated with DeepMAC for multitask learning which outperforms single task model.
• ensemble prediction using multiple detection models
• pseudo-label generation on test dataset to continue training the detection models

Our best-performing model achieved rank 3 on the test leaderboard. The details of the technique are discussed in our paper titled "Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022," which is available at IEEE ICIP 2022 at IEEE ICIP 2022 at https://ieeexplore.ieee.org/document/9897464. Please refer to the diagram below for a visual representation of the proposed techniques.

For qualitative analysis, you can see example bounding box and instance segmentation predictions of our final models on the given test set.

And the example of single model prediction and ensemble prediction.

Setup Instructions for Linux

Assuming Pytorch with GPU support is installed.

pip install mmcv-full==1.15.2 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html
pip install mmdet==2.25.0

Prediction

First download the trained models using the script,

cd models && bash get_model.sh

Then, run the following to get the ensembled predictions.

python predict_ensemble.py PATH_TO_IMAGES_FOLDER PATH_TO_MODEL_FOLDER --out SUBMISSION_JSON_FILE_NAME
# python predict_ensemble.py examples/ models/ --out pred_output.json

It may take up to 10 minutes to download the pretrained models from the repository. When making predictions and ensembling on the official test set, a single NVIDIA RTX2080Ti may take around 20 minutes to complete the task for approximately 1650 test images.

Results and Models

Individual models with their leaderboard scores, configs and checkpoints are shown in the table below.

Backbone	Architecture	Epochs	mIoU (Leaderboard)	Config	Checkpoint
HRNet	CascadeRCNN	10	0.927	config	ckpt
HRNet	HTC	10	0.928	config	ckpt
X-101-32x4d-dcnv2	HTC	10	0.928	config	ckpt
R-101-dcnv2	GFL	10	0.923	config	ckpt
R-101-dcnv2	TOOD	10	0.926	config	ckpt

Requirements

See requirements in requirements.txt including

• mmdetection

Citation

Cite an article as

Aung, Zaw Htet, Kittinan Srithaworn, and Titipat Achakulvisut. "Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022." In 2022 IEEE International Conference on Image Processing (ICIP), pp. 4273-4277. IEEE, 2022.

Or using Bibtex:

@inproceedings{aung2022multitask,
  title={Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022},
  author={Aung, Zaw Htet and Srithaworn, Kittinan and Achakulvisut, Titipat},
  booktitle={2022 IEEE International Conference on Image Processing (ICIP)},
  pages={4273--4277},
  year={2022},
  organization={IEEE}
}

Authors

• Zaw Htet Aung, Department of Biomedical Engineering, Mahidol University, Thailand
• Kittinan Srithaworn, Looloo technology, Thailand
• Titipat Achakulvisut, Department of Biomedical Engineering, Mahidol University, Thailand