hrlblab / imageseperation

ImageSeperation

Repository contains Python implementation of ...

• Training and testing code for sub-figure detector with sideloss
• cross validation on sub-figure detector, (results are essembled using weighted-box-fusion)

Installation

pip install -r requirements.txt

Data access

You need to contact the organizers of the task (https://www.imageclef.org/2016/medical) and ask for licensing the dataset.

Pre-trained model

To be released

Usage examples

Sub-figure detection

$ python detect.py --weights ./detector.pt --source ./image_dir --hide-labels(optional) --hide-conf(optional)

Subfigure crop

Train

$ python train.py --epochs 100 --batch-size 32 --data imageCLEF.yaml --weights yolov5s.pt --single-cls --sideloss

$ python train_cross_val.py --epochs 100 --batch-size 32 --data imageCLEF_cross_val.yaml --weights yolov5s.pt --single-cls --sideloss

Test

python test.py --batch 32 --data test.yaml --weight best.pt --single-cls --save-txt --save-conf

Model ensemble

test_merge.py

Citation

If you find this repository useful in your research, please cite:

Yao, T., Qu, C., Liu, Q., Deng, R., Tian, Y., Xu, J., Jha, A., Bao, S., Zhao, M., Fogo, A.B. and Landman, B.A., 2021. Compound Figure Separation of Biomedical Images with Side Loss. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections (pp. 173-183). Springer, Cham.

imageseperation's Issues

Location of the image enhancement code in the project

Your article describes an image enhancement method, can you tell me where this code is located?-------------------------------------

------------- we introduce an intra-class image augmentation method to simulate hard cases;

Any chance to get the pretrained weights for the model from your paper?

Hello! I was going through your paper Compound Figure Separation of Biomedical Images: Mining Large Datasets for Self-supervised Learning and wanted to try out this model but realized that you're not providing the pretrained weights with which you got the results mentioned in the paper and instead it is expected we finetune the model using your training process starting off from the YOLOv5 checkpoint.

While this is not really a deal breaker it means that (1) I need to get access to the data, and (2) wait for the training to finish.

So I wanted to ask if there's any chance you could somehow share the pretrained checkpoints that you used? 🤗

Thanks a lot!