InDiD: Instant Disorder Detection via Representation Learning

For sequential data, change points are moments of abrupt regime switches. Such changes appear in different scenarios, including complex video surveillance, and we need to detect them as fast as possible. Classic approaches for change point detection (CPD) perform poorly for semi-structured sequential data because of the absence of adequate data representation learning procedure. We propose a principled loss function that approximates classic rigorous solutions but is differentiable and makes possible representation learning. This loss function balances change detection delay and time to false alarm to provide a successful model for CPD. In experiments, we consider simple series and more complex real-world image sequences and videos with change points. For more complex problems, we show that we need more meaningful representations tailored for the specificity of the CPD task. Taking this into account, the proposed approach InDiD improves baseline results of CPD for various data types. For explosion detection, F1 score for our method is $0.54$ compared to baseline scores $0.46$ and $0.30$.