EthioDance‐ViT: A Video Vision Transformer for Classifying Ethiopian Traditional Dances

ABSTRACT Ethiopia is a country of exceptional cultural diversity, with more than 80 ethno‐linguistic groups, each of which preserves distinctive traditional dance forms. Automatic classification of short video clips into these regional dance traditions supports archival indexing, digital preservation, educational technology and the global movement to safeguard intangible cultural heritage. In this paper, we present EthioDance‐ViT, a Video Vision Transformer approach for classifying Ethiopian traditional dances into five major regional categories: Amhara, Oromo, Woliyta, Gurageya and Tigriga. We fine‐tune the TimeSformer model pretrained on Kinetics‐400 using a curated corpus of 18,060 short clips. A central methodological contribution is a content‐based perceptual‐hash grouping of clips that yields a group‐stratified train, validation and test split, which is designed to prevent near‐duplicate clips cut from the same source recording from inflating reported accuracy by appearing in both training and evaluation sets. Under the group‐stratified split, a straightforward cross‐entropy fine‐tune trained for up to 150 epochs achieves test accuracy and macro‐F1 across three random seeds. We further perform a systematic ablation of three popular class‐imbalance techniques (class‐balanced focal loss, weighted random sampling and video‐level MixUp) and find that none of them improves accuracy on this dataset. We complement the quantitative analysis with a class‐level visual atlas, an illustration of within‐class variation and example frames from each class. The resulting system and evaluation protocol offer a strong baseline for Ethiopian cultural‐video research.