umuteren/insightface_embeddings

Dataset Card for Insight Face Embeddings

Dataset Summary

This dataset contains face embeddings generated by the InsightFace model from the FaceData dataset. The embeddings are stored in .h5 files, with a total dataset size of 139GB. This dataset is useful for face recognition and verification tasks.

Dataset Structure

Data Instances

Each data instance consists of an embedding vector generated by the InsightFace model, along with the corresponding image. The embeddings and images are stored in HDF5 (.h5) files, where each file corresponds to a batch of processed images from the FaceData dataset.

Data Fields

Each .h5 file contains:

• /images/{image_name}: The resized and cropped image corresponding to the face.
• /embeddings/{image_name}: The embedding vector for the face.
• /filename_{image_name}: The original filename of the image, stored as an attribute.

Dataset Creation

Source Data

• Original Dataset: The embeddings are generated from the FaceData dataset, which contains face images used for training and evaluation.
• Model: InsightFace model was used to generate the embeddings.

Preprocessing

The preprocessing steps include:

1. Loading the images from the FaceData dataset.
2. Detecting faces using the InsightFace model.
3. Cropping and resizing the detected faces to a standard size.
4. Generating embeddings for the detected faces.
5. Storing the images and embeddings in HDF5 files.

Example

Here is an example of how to load and use the data:

import h5py
import numpy as np

def load_hdf5_data(hdf5_file):
    with h5py.File(hdf5_file, 'r') as h5f:
        images = {}
        embeddings = {}
        filenames = []

        for key in h5f['images'].keys():
            image_name = h5f.attrs[f'filename_{key}']
            img = np.array(h5f[f'images/{key}'])
            embedding = np.array(h5f[f'embeddings/{key}'])
            
            images[image_name] = img
            embeddings[image_name] = embedding
            filenames.append(image_name)

        return images, embeddings, filenames

# Example usage
hdf5_file = 'face_embeddings26.h5'
images, embeddings, filenames = load_hdf5_data(hdf5_file)

print("Loaded data from HDF5 file:")
print(f"Filename: {filenames[0]}")
print(f"Image shape: {images[filenames[0]].shape}")
print(f"Embedding shape: {embeddings[filenames[0]].shape}")