Efficient Video Object Classification Using Class Prior Adaptation Under Continuous Class-Balance Changes

Abstract

Video object classification is a fundamental task with broad applications such as anomaly detection in video streams and wildlife monitoring via camera traps. In practice, video classification is often decomposed into per-frame image classification. However, a major challenge arises from non-stationarity: class priors often vary over time, and a mismatch between training and test priors leads to performance degradation. Previous works have addressed this issue through runtime adaptation, where they exploit the temporal locality of the class occurrence and detect change points. One applies Bayesian filtering for adaptation against class prior change, while another pre-trains lightweight models under different priors and switches among them dynamically. Inspired by these approaches, we focus on class-balance changes over time. At a short timescale, class priors exhibit non-stationarity (e.g., different animals appear at different times of the day), while at a longer timescale, priors can show strong periodicity (e.g., seasonal animal activities). Unlike previous works based on change point detection, our approach leverages training data to predict priors in advance and dynamically adjusts the output of the original classifier at test time through a lightweight prior adaptation framework. Experiments on the wildlife ecological dataset show that our prior adaptation effectively improves classification accuracy under time-dependent class-balance change. This lightweight and effective adaptation scheme is particularly suited to edge computing environments with constrained resources.