All-SEAing Eye

Install mmdetection

Install mmdetection using mmdet_install.ipynb.

Download dataset

Download dataset from AI Hub 해상 객체 이미지

Rename image files and Generate annotation files

Organize the dataset directory as follows:

./mmdetection/data/
    test/
        images/
            20201112/
            ...
    train/
        images/
            20201216/
            ...
        labels/
            20201216/
            ...
    valid/
        images/
            20201224/
            ...
        labels/
            20201224/
            ...

Put anns_generate_files_rename.ipynb inside mmdetection/data and run it. It'll rename filenames from Korean to English (which is vital for mmdetection model to process images), and create annotation files test.json, train.json, and valid.json in COCO format.

Train and Inference

Place test.ipynb, and train.ipynb in mmdetection/tools and run train.ipynb to train, and test.ipynb to get inference.

Cascade R-CNN with ResNet-101 backbone is used.

With image size of 768x432, the inference speed is 22.5 FPS. mAP score is 0.226.

Inference results visualized: