TVFace

TVFace is a large-scale facial-recognition and clustering dataset built at NUST's MachVIS lab: 2,609,210 high-resolution face images of 28,955 unique individuals, extracted from broadcasts across 22 global television networks.

Because the source is live television, the data captures genuine real-world conditions (pose, lighting, expression, and aging over time) instead of the clean, frontal images most datasets rely on.

Facial-recognition research needs scale, diversity, and a realistic long tail. Most public datasets are small, curated, or demographically skewed. Without a benchmark that mirrors real distributions, it's hard to study unsupervised clustering, large-scale recognition, or demographic fairness honestly.

Faces were detected and clustered from television broadcasts into a natural long-tail distribution, from 10 to 21,983 images per identity, at 224×224 resolution. Each identity carries probabilistic demographic annotations: age, gender, ethnicity (7 groups), expression (7 categories), and head pose.

The dataset ships with a PyTorch-compatible loader and an ethics-first usage framework: research-only, non-commercial, with explicit privacy and bias-mitigation requirements.

A research resource for unsupervised facial clustering, large-scale recognition, and demographic-fairness analysis. Built at MachVIS and published in Springer's Pattern Analysis and Applications, under a Creative Commons non-commercial license.

It now anchors the TVFace Challenge 2026, a public clustering-and-retrieval competition built on the dataset.

01Real-world distributions beat clean ones for honest benchmarks.

02A long tail is a feature, not a defect, when you study recognition.

03Scale and ethics have to ship together.