This repository is the implementation of RSNA Pneumonia Detection Challenge. The project compares and analyzes the performance of object detection models such as Faster/Cascade RCNN, Retinanet and Swin Transformer on RSNA dataset. Also, some ensemble and post processing methods are used in this project.

In this project, framework MMDetection is used. The package should be installed first to implement further analysis. Refer to MMDetection official installation for more information.

# Create a conda virtual environment and activate it.
conda create -n openmmlab python=3.7 -y
conda activate openmmlab

# Install PyTorch and torchvision following the official instructions
conda install pytorch torchvision -c pytorch

# Install MMDetection
pip install openmim
mim install mmdet

The dataset used in this project could be downloaded using the folloing kaggle command or via this link. To transform labels into Coco dataset style, run the following command:

kaggle competitions download -c rsna-pneumonia-detection-challenge
cd utils
python coco_formatter.py

Import the customized model configurations into MMDetection configs directory first and then run the training script.

mv <model name>/<configs.py> <path of mmdetection>/config/<model name>

cd mmdetection
python tools/train.py configs/<model name>/<configs.py>

The weights of the models above is saved in Drive: Trained Models

To infer the testing images using trained model, run this to generate bbox file in .json format

cd mmdetection
./tools/dist_test.sh \
    /work_dirs/<model name>/<model configs.py> \
    /work_dirs/<mode name>/<model weights.pth> \
    1 \
    --format-only \
    --options "jsonfile_prefix=./<inference filename>"

To post-process the result and submit the answer to kaggle, run submission.py, ensemble.py and resize_submission.py in order.

TBD