Deities-25

The dataset comprises of a comprehensive collection of 8,239 images showcasing diverse forms and iconographies of 25 Indic deities. This dataset is a unique blend of manually curated and web-scraped visuals, providing a valuable resource for the computer vision community interested in exploring the artistic and cultural expressions embedded in the visual representation of deities.

Supported Tasks

• image-classification: The goal of this task is to classify a given image of a deity into one of 25 classes.

Uses

Direct Use

• Cultural Awareness: Raise awareness about the rich cultural heritage of the Indian subcontinent by incorporating these diverse depictions of Indic deities into educational materials.
• Research and Preservation: Contribute to academic research in the fields of art history, cultural studies, and anthropology. The dataset serves as a valuable resource for preserving and studying the visual representations of revered figures.
• Deep learning research: Offers exciting opportunities for multi-label classification tasks. However, a challenge in this domain is dealing with inter-class similarity, where images from different categories share common features.

Source Data

Social media posts, smartphone camera captures, images generated using diffusion methods.

Data Collection and Processing

We carefully selected diverse images for the dataset and used the cleanvision library from cleanlab to remove images with issues. A custom Python script helped organize the data effectively. When it came to training our model, we relied on torchvision transforms to prepare our dataset for training.

Dataset Structure

DatasetDict({
    train: Dataset({
        features: ['image', 'label'],
        num_rows: 6583
    })
    validation: Dataset({
        features: ['image', 'label'],
        num_rows: 1656
    })
})

Dataset Splits

This dataset is split into a train and validation split. The split sizes are as follow:

Split name	Num samples
train	6583
valid	1656

Bias, Risks, and Limitations

• Bias - The dataset primarily represents Indic deities, potentially introducing a cultural bias. Efforts were made to include diverse forms, but the dataset may not fully encapsulate the breadth of artistic expressions across different Indic cultures.
• Risks - Images of deities can be open to various interpretations. The dataset may not capture nuanced meanings, leading to potential misinterpretations by users.