Do Outliers Matter?

Training sample quality impacts deep learning model performance. While studies in the literature explored the association of outlier samples to model performance in modalities like text or images in the NLP and computer vision domains, it is relatively underexplored in the domain of video classification. Researchers focused on anomaly detection or theoretical bounding of outliers towards video classification, but explicit, systematic empirical studies of the impacts of these outliers on video classification modeling are still yet to be explored.

To bridge this gap, this work systematically analyzes the impacts of outliers, specifically in-distribution outliers, on video classification performance and shows that reducing outliers from training can improve video classification results.