This work aims to achieve high-precision 3D face reconstruction from a sequence of uncalibrated multi-view images acquired directly by the user. The purpose is to compute optical measurements on the face. We therefore focus on reconstructing the face in a neutral expression using a learning-based approach, initialized by a model, which allow for predictions beyond the model’s statistics. We tailor an algorithmic pipeline to directly estimate the 3D face model using a multi-view stereovision method. We demonstrate that using a 3DCNN to predict surface proximity enables the network to generalise to unseen and uncalibrated datasets while training on calibrated data. We enhance performance through three proposed improvements: using a dense algorithm to retrieve the initial 3D face, applying image pre-processing, and introducing data augmentation. To demonstrate this behaviour, our method is evaluated on three datasets of uncalibrated images. Beyond generalising to unseen datasets, our approach outperforms state-of-the-art methods in terms of accuracy while maintaining a fast computation time.