Reputation: 23
The task is to train ready open source neural network for open-set face recognition through the database of 10k users with each of them shot for 2-3 images (filmed from a little different angles)
What level of accuracy could be expected?
Upvotes: 1
Views: 2304
Reputation: 3408
The dataset is somewhat larger than "Labeled Faces in the wild" (around 5000 fases, with 1-240 samples for each face). For that dataset, neural networks have been able to pass 90%. While it depends on the quality of photos, how different the people look, etc., for large numbers 90% is a reasonable estimate. Of course, your results can be worse if the network is too small and images are not preprocessed well (de-rotation of faces is common in good classifiers).
This paper is a survey on systems that have been tested on the above dataset.
Upvotes: 1