Evaluation Metric "Classification Error" on RaFD dataset?

[Bidur-Khanal]

I am unable to reproduce the reported classification error. In the paper, the reported value is 2.12 and 0.45 for translated and real images, respectively, of RaFD dataset.

I implemented a similar face classifier architecture that is described in the paper. I obtained a classification error of 0.45 (same) on real images. However, the error on the translated images is large (almost 8.1).
Are you using extra preprocessing techniques on faces (example face aligning) or training for a longer period to obtain low classification errors?

I need this metric to establish a baseline for a new project I am working on. I would appreciate your help