Face Search at Scale.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Searching for specific individuals across vast numbers of unconstrained images on social media is a significant challenge.
• Existing face recognition methods struggle with the scale and variability of real-world image data.

Purpose of the Study:

• To develop a scalable and accurate face search system for large image datasets.
• To improve face search performance by combining deep learning features with commercial matching algorithms.

Main Methods:

• A cascaded framework was employed, initially filtering images using convolutional neural network (CNN) features.
• Top-k candidates were re-ranked by fusing CNN-derived similarities with those from a commercial off-the-shelf (COTS) matcher.
• The system was evaluated on an 80 million image gallery and standard benchmarks like LFW and IJB-A.

Main Results:

• The fused approach achieved 99.5% True Accept Rate (TAR) at 0.01% False Accept Rate (FAR), outperforming individual methods on a mugshot dataset.
• On unconstrained benchmarks, the system demonstrated competitive performance, including 82.2% rank-1 retrieval on IJB-A.
• The system successfully identified a person of interest in a large gallery within seconds.

Conclusions:

• Combining deep features with COTS matchers offers complementary information, significantly boosting face search accuracy.
• The proposed cascaded system provides an effective balance between accuracy and scalability for massive image galleries.
• This approach is highly promising for real-world applications like forensic investigations and content moderation.