Facial expression recognition (FER) has gained significant attention due to its applications in fields such as healthcare, education, and social robotics. However, its application to virtual reality headset users is challenging due to the occlusion generated by the device. This paper explores FER using only the lower third of the face, introducing a novel pooling method that transforms static image classification architectures into robust video classification models. We evaluated the method on CK+ and Oulu-CASIA datasets, comparing its performance with state-of-the-art video processing architectures. The models studied in this paper outperform the state-of-the-art achieving up to 93.3% classification accuracy in the CK+ dataset with 70% occlusion. The results show that the evaluated models with the proposed pooling achieve competitive accuracy and exhibit robustness against temporal variations. The proposed pooling method can potentially be applied to any existing image classification deep learning architecture without increasing the number of trainable parameters.