Lightweight dynamic large convolution model for real-time human pose estimation
Zongyou Liu, Shilan Liu, Ziyuan Liu, Hongtao Wang, Quanxin Jin · Third International Conference on Computer Science and Communication Technology (ICCSCT 2022) · 2022
Although the lightweight Openpose model can perform real-time human pose estimation on the CPU, it also has the following problems: (1) The ReLU activation function destroys the narrow channel information; (2) Although the model has high inference efficiency, it also has weak feature extraction ability; (3) The model has a small effective receptive field. In order to improve the above problems, we adjusted the structure of the lightweight Openpose model, replacing the backbone of the lightweight Openpose with the backbone of MobileNetV2 to improve the destruction of the narrow channel information by the ReLU activation function. Replacing part of the backbone’s convolutional layers with dynamic depth wise separable large convolutional layers expands the effective receptive field of the model, improves feature extraction and slightly increases the computational cost. Our Lightweight dynamic large convolution model achieves an Average Precision (AP) value of 0.40 on the COCO2017 validation set.