GSFNet: A Lightweight Global Semantic Fusion Pose Estimation Network
Hailan Xu, Lei Liu, Cong Chen, Shengzhao Hao, Shiyu Li, Zhi Liu · 2023
Multi-person pose estimation is an important research direction in the field of computer vision. Traditional top-down and bottom-up methods divide this task into two stages for pose estimation, leading to additional computational overhead. Therefore, this paper proposes an end-to-end lightweight multiperson pose estimation network called GSFNet based on YOLO-Pose. This network maintains lightweightness while improving the accuracy and speed of pose detection. To address issues such as redundant convolutional feature maps and low accuracy in detecting small-sized objects in the network, we first construct a C3P module using Partial Convolution (PConv) as the feature extraction module, which effectively alleviates the redundancy of network feature maps while reducing the number of parameters and computations. Then, we propose gsBiFPN (global semantic bidirectional feature pyramid network) for multi-scale feature fusion to enhance the detection capability of small-sized objects, thereby improving the overall performance of the model. Experiments demonstrate that compared to the original YOLO-Pose baseline method, this approach achieves a 7.4% increase in Average Precision (AP) accuracy on the COCO 2017 validation dataset while reducing the network parameters and computational load.