The paper introduces the MegaFace benchmark, a dataset and challenge designed to evaluate and advance face recognition algorithms at scale. The MegaFace dataset includes one million photos of over 690,000 unique individuals, collected from Yahoo's Flickr database. The challenge assesses algorithms' performance with increasing numbers of "distractors" (up to one million people not in the test set) and evaluates both identification and verification tasks. Key findings include: 1. **Performance Scaling**: Algorithms that achieve high performance on smaller benchmarks (e.g., LFW with 10 distractors) perform significantly worse with one million distractors, with rates dropping from above 95% to 35-75%. 2. **Training Data Size**: Algorithms trained on larger datasets (e.g., FaceNet trained on 500 million photos) generally perform better at scale, though there are exceptions (e.g., FaceN trained on 18 million photos). 3. **Age and Pose Effects**: Recognition performance is more challenging with age differences and varied poses, especially at larger scales. 4. **Age Invariance**: Children (below age 20) are more difficult to recognize than adults, and larger age gaps between gallery and probe faces are harder to match. The paper also discusses the creation of the MegaFace dataset, the evaluation protocols, and presents results from state-of-the-art algorithms. The MegaFace benchmark is publicly available to encourage further research and development in face recognition.The paper introduces the MegaFace benchmark, a dataset and challenge designed to evaluate and advance face recognition algorithms at scale. The MegaFace dataset includes one million photos of over 690,000 unique individuals, collected from Yahoo's Flickr database. The challenge assesses algorithms' performance with increasing numbers of "distractors" (up to one million people not in the test set) and evaluates both identification and verification tasks. Key findings include: 1. **Performance Scaling**: Algorithms that achieve high performance on smaller benchmarks (e.g., LFW with 10 distractors) perform significantly worse with one million distractors, with rates dropping from above 95% to 35-75%. 2. **Training Data Size**: Algorithms trained on larger datasets (e.g., FaceNet trained on 500 million photos) generally perform better at scale, though there are exceptions (e.g., FaceN trained on 18 million photos). 3. **Age and Pose Effects**: Recognition performance is more challenging with age differences and varied poses, especially at larger scales. 4. **Age Invariance**: Children (below age 20) are more difficult to recognize than adults, and larger age gaps between gallery and probe faces are harder to match. The paper also discusses the creation of the MegaFace dataset, the evaluation protocols, and presents results from state-of-the-art algorithms. The MegaFace benchmark is publicly available to encourage further research and development in face recognition.