Mamba-MSQNet: A Fast and Efficient Model for Animal Action Recognition

Animal action recognition is crucial for assessing animal well-being in agriculture and environmental monitoring. Recent advancements in this field rely on computer vision technologies. However, many current applications are restricted to recognizing actions within a single animal species or a limited set of actions, resulting in highly specific models and lacking generality. When addressing a broader range of actions and species, transformer models are typically required, which demand significant computational and processing power, potentially limiting their practical use. In this work, we introduce a deep learning model based on selective state spaces designed to reduce the computational cost of MSQNet, the current state-of-the-art model for action recognition in the Animal Kingdom dataset. Our approach achieves superior results with fewer parameters and lower FLOPs, thereby enhancing efficiency without compromising performance. Code available on https://github.com/edofazza/mamba-msqnet.