Interaction Relational Network for Mutual Action Recognition

Person-person mutual action recognition (also referred to as interaction recognition) is an important research branch of human activity analysis. Current solutions in the field – mainly dominated by CNNs, GCNs and LSTMs often consist complicated architectures mechanisms embed relationships between two persons on architecture itself, ensure patterns can be properly learned. Our main contribution with this work proposing a simpler yet very powerful architecture, named Interaction Relational Network, which utilizes minimal prior knowledge about structure body. We drive network identify itself how relate body parts from individuals interacting. In order better represent interaction, we define different relationships, leading specialized models for each. These multiple relationship will then fused into single special leverage both streams information further enhancing relational reasoning capability. Furthermore structured pair-wise operations extract meaningful extra each pair joints distance motion. Ultimately, coupling LSTM, our IRN capable paramount sequential reasoning. extensions made also valuable other problems that require sophisticated solution able achieve state-of-the-art performance traditional datasets SBU UT, actions large-scale dataset NTU RGB+D. Furthermore, it obtains competitive RGB+D 120 interactions subset.