ExpT: Online Action Detection via Exemplar-Enhanced Transformer for Secondary School Experimental Evaluation

Secondary school experimental evaluation is an essential component of secondary school science education. However, it faces several challenges, including obstacles to precise assessment within limited time and the presence of inconsistent evaluation criteria. Hence, it has become imperative to explore and harness artificial intelligence technology to improve secondary school experimental evaluation. Yet existing applicable online action detection (OAD) algorithms are hindered by limitation to historical context and inefficiency, leading to setbacks in realistic experimental evaluations. Based on this, we present Exemplar-enhanced Transformer (ExpT), a real-time mechanism for online action detection that more accurately and efficiently assesses the experiments conducted by students. By leveraging exemplars through temporal cross attention, the ExpT model provides complementary guidance for modeling temporal dependencies, along with the reduction of excessive attention. We evaluate ExpT on two realistic chemistry experiment datasets for online action detection, and it significantly outperforms all existing methods.