基于多样卷积单元高效人体姿态估计

Efficient Human Pose Estimation Based on Diverse Convolutional Units

中文关键词:  高分辨率  人体姿态估计  卷积神经网络  卷积单元

英文关键词:high-resolution  human pose estimation  convolutional neural networks  convolution unit

中文摘要:

      设计兼具准确率和轻量化的人体姿态估计网络模型成为了人机交互领域的迫切需求。为了满足这一需求，结合HRNet模型的高分辨率设计模式，提出了一种多样化高效卷积单元的高分辨率网络模型DU-HRNet。为了探索并增强来自不同感受野大小层的多尺度信息，鼓励卷积层间信息更加多样化，模型允许并行分支中的每一分支拥有不同类型的高效卷积单元。为了改善模型的非线性，在高效卷积单元中使用通道注意力ECANet。在MS COCO关键点检测数据集和MPII数据集中验证了模型的有效性。模型在参数量等于7.6 M、GFLOPs为2.66，没有经过任何后期处理的条件下，在COCO val2017数据集上达到了71.1 mAP（mean Average Precision,平均精度均值）分数，在COCO test-dev2017数据集上达到71.8 mAP分数。通过消融实验验证了模型整体和组成部分的有效性。

英文摘要:

      It is an urgent demand in the field of human-computer interaction to design a network model for human posture estimation with both accuracy and lightweight.In order to meet this need,the high-resolution network model DU-HRNet with diverse high-efficiency convolutional units is proposed by combining the high-resolution design model of HRNet network model.In order to explore and enhance the multi-scale information from different receptive field size layers and to encourage more diverse information among convolutional layers,the model allows each branch of parallel branches to have different types of high-efficiency convolutional units.In order to improve the nonlinearity of the model,channel attention ECANet is used in the efficient convolution units.The effectiveness of the model is validated on the MS COCO keypoint detection dataset and the MPII dataset.The model achieves 71.1 mAP scores on the COCO val2017 dataset and 71.8 mAP scores on the COCO test-dev2017 dataset with the number of parameters equal to 7.6 M,GFLOPs of 2.66,and without any post-processing.The validity of the whole model and its components is verified by ablation experiments.