README.md

metadata

title: MobileCLIP Image Embedding Extractor
emoji: 🚀
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
short_description: Extract image embeddings using MobileCLIP-S2 model
tags:
  - computer-vision
  - image-embeddings
  - mobileclip
  - clip
  - zero-shot
  - image-classification

MobileCLIP Image Embedding Extractor

This Gradio app allows you to extract image embedding vectors using the MobileCLIP-S2 model.

Features

• Fast Inference: Uses MobileCLIP-S2 optimized for mobile deployment
• High Quality Embeddings: 512-dimensional normalized embedding vectors
• Easy to Use: Simple drag-and-drop interface
• Multiple Formats: Supports JPG, PNG, GIF, BMP, TIFF

About MobileCLIP

MobileCLIP is a family of efficient image-text models that achieve excellent performance while being significantly faster and smaller than comparable models. MobileCLIP-S2 provides:

• 2.3x faster inference than SigLIP ViT-B/16
• 2.1x smaller model size
• Better average zero-shot performance
• Trained with 3x fewer samples

Usage

1. Upload an image using the interface
2. Click "Extract Embedding" or the embedding will be automatically extracted
3. View the summary and download the embedding vector in JSON format

Model Details

• Architecture: MobileCLIP-S2
• Training Dataset: DataCompDR
• Embedding Dimension: 512
• Normalization: L2 normalized embeddings

Citation

@inproceedings{vasu2024mobileclip,
  title={MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training},
  author={Vasu, Pavan Kumar Anasosalu and Pouransari, Hadi and Faghri, Fartash and Vemulapalli, Raviteja and Tuzel, Oncel},
  booktitle={CVPR},
  year={2024}
}