Chess Dataset
A simple and elegant image dataset for beginners
Mission of the project
To prepare a machine learning dataset similar to MNIST using image transformation techniques for beginners to perform image classification.
Dataset

Figure-1.1: Dataset Statistics

Figure-1.2: Sample of Dataset-1
Here basic image transformations were applied to raw images. It occupies 1GB of memory space.

Figure-1.3: Sample of Dataset-2
Some of the images of this dataset are blurred, grey scaled, and have undergone histogram equalization. It occupies 770MB of memory space.
Raw Data

Figure-1.4: Raw Data Statistics(1)
The chess pieces were captured at 7 vertical angles as mentioned above.

Figure-1.5: Raw Data Statistics(2)
At each vertical angle, the images of chess pieces were captured in a 360° view .
Data Download
Verification
The usefulness of the dataset has been verified using classification algorithms like SVM, KNN, and CNN.
Results

Figure-1.6: SVM pairwise classification accuracies
SVM
The pairwise classification was performed with SVM. The data was split in a ratio of 4:1. The SVM classifier was trained on all images. The accuracy for each pair can be found in figure 1.6.

Figure-1.7: Accuracies of KNN (1.0) and KNN(2.0) models.
KNN
The train and test data were split in the ratio of 4:1. Initially, the model was trained with 50 images per chess piece (1.0), later the image count was increased to 500 images per chess piece (2.0). The classification was performed for 6 chess pieces using the KNN algorithm. The resultant accuracy can be found in figure-1.7.

Figure-1.8: CNN accuracies.
CNN
1000 images for each class were taken for training the CNN model. The dataset was split into 4:1 ratio. The batch size is 100, the epochs are 1000 and the activation function used was ReLu. the resulting accuracy can be found in figure-1.8.