YOWOv3: An Efficient and Generalized Framework for Spatiotemporal Action Detection

In this paper, we propose a significant enhancement of YOWOv2, resulting in a lightweight and efficient action detection and recognition model referred to as YOWOv3. This model utilizes recently proposed architectures to reduce computational costs while integrating an enhanced label assignment mechanism to preserve model performance, thereby compensating for the decreased complexity of the architecture. Experiments were conducted on two popular datasets, UCF101-24 and AVAv2.2, to evaluate the proposed YOWOv3. YOWOv3 illustrates competitive results, obtains an mAP of 20.93% on AVAv2.2, and 88.64% on UCF101-24, by utilizing just 54.7M parameters and 37.2 GFLOPs compared to benchmark YOWOv2 with 109.7M parameters and 53.6 GFLOPs, as well as an mAP of 85.2% and 20.3% on UCF101-24 and AVAv2.2, respectively. The code is publicly available at: <ext-link ext-link-type="uri" xlink:href="https://github.com/AakiraOtok/YOWOv3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/AakiraOtok/YOWOv3</ext-link>.