A collection of RGB images of vinyl album covers labeled with one of three music genres: electronic, rock, and hiphop. The dataset is designed for image-based genre classification and model interpretability studies. It originates from a flat “Training” directory paired with a CSV of labels, then is split into train/validation/test and reorganized into a class-structured folder hierarchy for model training with Keras/TensorFlow. Images are typically resized to 300×300 and rescaled to [0, 1].
Use cases
• Multi-class image classification (3 classes)
• Transfer learning with CNN backbones (e.g., ResNet152V2)
• Interpretability research via Concept Whitening or related techniques
• Educational benchmarks for image data pipelines (splitting, directory structuring, generators)
Dataset structure
• Original source
• Images directory: For thesis/Training/
• Labels CSV: For thesis/training_labels.csv
◦ Columns:
◦ name: image filename (e.g., cover123.jpg)
◦ category: string label ∈ {el…
