Hey, I'm the author of this work and the maintainer of DeepFace.
In this paper, we ensemble the top 5 DeepFace backbones and feed them into a GBM model instead of relying on a single embedding model.
On LFW, this surpasses FaceNet512’s 98.4% accuracy and the reported human-level accuracy of 97.5%, achieving 99.1%.
We deliberately avoid retraining on LFW to prevent benchmark overfitting. The base models are trained on large-scale datasets, and we only learn a boosting layer on top of their similarity scores.
In this paper, we ensemble the top 5 DeepFace backbones and feed them into a GBM model instead of relying on a single embedding model.
On LFW, this surpasses FaceNet512’s 98.4% accuracy and the reported human-level accuracy of 97.5%, achieving 99.1%.
We deliberately avoid retraining on LFW to prevent benchmark overfitting. The base models are trained on large-scale datasets, and we only learn a boosting layer on top of their similarity scores.