Labeled Faces in the Wild Subset with 72 Facial Attributes

A subset of 400 grayscale face images from the Labeled Faces in the Wild collection, preprocessed for tensor-on-tensor regression. Each 30x30 pixel image is paired with a 72-dimensional vector of continuous facial attributes. Fabio Centofanti published this version in March 2026 for use in a study on robust tensor regression.

Use Cases

• Predict 72 facial attributes like gender and expression from 30x30 grayscale image tensors.
• Train robust regression models to map image pixel data to continuous attribute vectors.
• Benchmark tensor-on-tensor regression methods using a known face dataset with pre-extracted features.
• Analyze model performance on facial data with variability in lighting and pose.

Strengths

• Each image is annotated with 72 continuous facial attributes.
• Images are standardized to a consistent 30x30 grayscale resolution.
• Dataset is focused on 400 distinct individuals, reducing identity redundancy.

Limitations

• Very small subset of 400 images, limiting statistical power for broad generalization.
• Images are heavily downsampled to 30x30 pixels, losing high-resolution detail.
• Dataset is derived from web images, which may introduce selection bias in demographics.

Provenance

Source

Subset of the public Labeled Faces in the Wild (LFW) dataset.

Collection Method

Images gathered from the web, converted to grayscale, downsampled, and paired with precomputed attribute vectors.

Time Range

null

Freshness

Last updated March 2026.

Geography

null