Playing Cards

This repository implements Single Shot MultiBox Detector (SSD) on a costume dataset. The objective is to detect the cards' club and number.

The SSD network is implemented with Pytorch library.

Files

The files in the repository are

• cards_class.py: Dataset class
• detect.ipynb / detect.py: Use a model to dtect objects in a given image
• eval.ipynb / eval.py: Evaluate the model using Mean Average Precision (mAP)
• model.py: SSD Neural Network using VGG-16 architecture and pretrained model for clasification
• relabeling.py: Relabeling if more images are added
• split_labels.py: Remove not wanted labels
• train.ipynb / train.py: Train the SSD Neural Network with costum dataset
• utils.py: Functions used by other files

Dataset

For the dataset, standard Bicicle Playing Cards where used for training and validation sets. Images where taken from different lightning, rotation and size conditions.

The images are stored in data/images, their anotations are in data/txt_cards and the labels are in data/general_labels.

Model

The model's output is stored in data/models.

References

I used as a reference sgrvind's github repository a-PyTorch-Tutorial-to-Object-Detection in which the concepts behind this neural network is very well explained.