The facial expression recognition (FER) system has always been in trend and majorly when it comes to the FER from the video clips, then it becomes a crucial task to do. The differences seen in the visual descriptor and the emotions seen on the face need to be filled in the FER from videos. Here in our proposed work aggregation is done between the spatial and temporal convolutional features available in the whole video to recognize facial expressions in any video. We are using here 15-15 both spatial and temporal streams. Every stream, i.e. spatial, corresponds with the temporal flow, which creates a layer of aggregation for end-to-end FER system training from the video. This training gives a better representation ofthe video and avoids overfitting from the limitation of available datasets. We have found that the proposed approach is best for aggregating the video dataset's spatial-temporal features compared to other available methods. The dataset we have used are RML, MMI, BAUM-1s, FER-2013 and the results obtained after applying the proposed approach on this datasets are satisfactory

Video - Face Expression Recognition, Feature Aggregation, CNN, Spatial-Temporal Features, Max-Pooling Laye