Most existing methods pay much attention to how to improve the accuracy of human pose estimation results. They usually ignore what the size of their model is. However, besides accuracy, real-time and speed are also important. In this paper, a new module named Densely Connected Residual Module is presented to effectively decrease the number of parameters in our network. We introduce our module into the backbone of High-Resolution Net. In addition, we change direct addition fusion into pyramid fusion at the end of the network. No need for ImageNet pre-training sharply decreases the total time of our training processes. We do our experiments over two benchmark datasets: the COCO keypoint detection dataset and the MPII Human Pose dataset. As a result, we achieve a decrease on number of parameters and calculated amount, respectively by around 72% and 14%, making our network more lightweight than High-Resolution Net. During testing process, our model can predict an image at a speed of 25 ms per image, which also achieves real-time fundamentally. The code has been available at https://github.com/consistent1997/LDCRN.

现有的大多数方法都非常关注如何提高人体姿态估计结果的准确性。他们通常忽略模型的大小。但是，除了准确性外，实时性和速度也很重要。在本文中，提出了一个名为“密集连接残差模块”的新模块，以有效减少我们网络中的参数数量。我们将模块引入高分辨率网络的骨干中。此外，我们在网络末端将直接加法融合更改为金字塔融合。无需对ImageNet进行预培训，就可以大大减少培训过程的总时间。我们在两个基准数据集上进行了实验：COCO关键点检测数据集和MPII Human Pose数据集。结果，我们减少了参数数量和计算量，分别增加了约72％和14％，使我们的网络比高分辨率网络更轻便。在测试过程中，我们的模型可以以每张图像25毫秒的速度预测一张图像，这也从根本上实现了实时。该代码已在https://github.com/consistent1997/LDCRN上提供。