EIDORA/DINO_ViTS16_IN1k

DINO ViT-S/16 Image Medium

DINO ViT-S/16 is a self-supervised Vision Transformer adapted for image embeddings in EIDORA. It produces 384-dimensional representations for exploring visual similarity in image collections.

Training Data and Checkpoint Provenance

• Architecture: Vision Transformer Small, patch size 16 (ViT-S/16).
• Original checkpoint: facebook/dino-vits16.
• Pinned revision: abe3b354cb6a9b6f146096b14a4a9d7eecbcb4bd.
• Training data: ImageNet-1K.
• Training method: Self-supervised DINO training, without ImageNet class labels as supervision.
• Source: Meta AI DINO checkpoint, distributed through Hugging Face.

DINO learns representations from images rather than paired image-text descriptions. The EIDORA package uses the pretrained visual model to produce normalized image embeddings.

Model Overview

• Model: DINO ViT-S/16 Image Medium
• EIDORA identifier: eidora/image-dino-vits16-medium
• Modality: Image
• Embedding dimensions: 384
• Embedding normalization: L2-normalized
• Similarity measure: Cosine similarity
• Package schema: Version 3
• Model version: 0.1.0

EIDORA Use and Supported Modalities

DINO is available as an image embedding model in EIDORA.

Researchers can use it to:

• Compare images based on learned visual representations.
• Explore visually related items in a collection.
• Investigate clusters and similarity patterns.
• Compare self-supervised DINO embeddings with CNN-based or CLIP embeddings.

Supported modality: Images only.

DINO does not provide text embeddings or a shared image-text embedding space.

Input and Preprocessing

EIDORA performs image preparation before passing the tensor to the ONNX model.

The default preparation is:

1. Convert the image to RGB.
2. Resize directly to 224 × 224 using the configured stretch method.
3. Use bilinear interpolation.
4. Rescale pixel intensities by 0.00392156862745098.
5. Normalize the RGB channels using the configured mean and standard deviation.
6. Arrange the tensor in NCHW format.

Input size: 224 × 224 pixels.

Channel mean: [0.485, 0.456, 0.406]

Channel standard deviation: [0.229, 0.224, 0.225]

Tensor layout: NCHW

Important: Channel normalization is part of EIDORA's preprocessing contract for this schema-v3 package. It must not be applied a second time inside a calling application.

The exact default settings and available controls are defined in config.yaml. Alternative settings should not be assumed equivalent to the documented defaults.

Output Embeddings

The package returns an image embedding with the following properties:

• Data type: float32
• Shape: [batch, 384]
• Normalization: L2-normalized
• Recommended similarity: Cosine similarity

The embedding represents learned image features. It is not a prediction of historical meaning or cultural significance.

Runtime and Compatibility

This package uses ONNX Runtime and EIDORA's schema-v3 configuration.

The package declares CPUExecutionProvider for CPU execution. The available execution providers depend on the user's runtime environment.

EIDORA uses config.yaml for image preparation, model inputs, and embedding output information.

Limitations

• DINO was pretrained on ImageNet-1K, not specifically on historical or cultural collections.
• Visual similarity does not necessarily mean historical, cultural or semantic similarity.
• This model does not accept text as a second modality.
• Other preparation settings exposed by EIDORA are not guaranteed to preserve the same representations as the documented default.
• Results should be assessed in relation to the dataset and research question.

Research Context and References

Main paper

Caron, M. et al. (2021). Emerging Properties in Self-Supervised Vision Transformers. ICCV.

Paper: https://arxiv.org/abs/2104.14294

Abstract summary: The paper studies self-supervised learning with Vision Transformers. It shows that DINO can learn useful visual representations without supervised class labels, including features that capture aspects of image structure.

The research motivates using DINO for visual representation analysis. It does not establish that this particular EIDORA embedding is optimal for every research dataset.

Primary citation (BibTeX)

@inproceedings{caron2021emerging,
  title={Emerging Properties in Self-Supervised Vision Transformers},
  author={Caron, Mathilde and Touvron, Hugo and Misra, Ishan and Jegou, Herve and Mairal, Julien and Bojanowski, Piotr and Joulin, Armand},
  booktitle={ICCV},
  year={2021}
}

License and Attribution

The upstream facebook/dino-vits16 checkpoint is released under Apache License 2.0.

The original model is attributed to Meta AI and the authors of the DINO research paper.

EIDORA provides the ONNX conversion, embedding adaptation and package configuration. It does not claim authorship of the original model.

See the included LICENSE file.

Package Information

• Model repository: EIDORA/DINO_ViTS16_IN1k
• Package version: 0.1.0
• ONNX model: model.onnx
• EIDORA configuration: config.yaml
• Package schema: 3
• License identifier: apache-2.0
• Supported modality: Image
• Declared provider: CPU

The configuration contains the authoritative embedding, preparation, artifact-integrity and runtime declarations for this package.