Hangul-SSC

Single Shot Classifier for Hangul Classfication
Forked from ssd.pytorch by amdegroot

Environment

• CUDA toolkit 11.0
• Pytorch 1.7.0
• SciPy 1.5.3
• Numpy 1.19.2
• imgaug 0.4.0

Dataset

We used HIL-SERI, which is intersection of HIL and SERI95 dataset. HIL dataset is only accessible via EIRIC website, and you can get SERI95 dataset from HangulDB repository. We took only 128 classes on them, and took some preprocessing.

• Note: To get HIL dataset, you must wrote memorandum to only use dataset on reseaching purpose! By this reason, we do not provide dataset.

Training

• With plenty data, you can start your own training.
• First, you have to download the fc-reduced VGG-16 PyTorch base network weights at: https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
• After download, place weight file on root directory of this project.
• You now can train SSC model with train.py file. Check argument for detail.

Evaluation

• To evaluate trained model, you can use eval.py or eval_list.py.
• eval.py is for single model file. you have to put model file's directory.
• eval_list.py is for multiple model files. you have to put directory which contains model files.

Accuracy

On preprocessed SERI-95 dataset, SSC scored accuracy of 98.56%.