Support Vector Machines are a set of machine learning models that have great performance in several tasks as well as on image classification and object recognition. However, the proper choice of model's hyperparameters has a great influence on the outcomes and the general capacity performance. In this paper, we explore some different traditional auto-tuning processes to estimate σ hyper-parameter for SVMs Gaussian kernel. These processes are common and also implemented on standard software of data science languages. The paper considers some different situations on smile detection. The results are composed by simulation study, two benchmark image applications and a real video data application.