Enhancing Multi-modal Automatic Target Recognition using Out-of-Distribution Exploitation (MATRODE)

Automatic target recognition (ATR) has seen significant advancements through multi-modal data and deep learning techniques. However, offline-trained ATR models using in-distribution data (IDD) often suffer performance degradation when encountering out-of-distribution (OOD) data that exhibits characteristics different from the training dataset. This paper introduces MATRODE which leverages multi-modal data, including electro-optical (EO), infrared (IR), acoustic, passive RF (pRF), and radar, without relying on modality-specific encoders. Instead, it employs topological data analysis (TDA) to extract topological features in a training free process. Unimodal machine learning models are developed by incorporating Conformal Prediction to identify OOD modalities. Only IDD modalities are fused for ATR to ensure robust performance in the presence of OOD data. Experiments on the multi-modal AFRL/RI ESCAPE II dataset demonstrate that the proposed approach effectively detects OOD data while maintaining high ATR accuracy with multi-modal inputs.