基于Transformer的野生动物关键点检测

Wildlife Pose Estimation Based on Transformer

濒危野生动物需要得到保护,对野生动物进行姿态追踪是保护野生动物的一种可行性方法。关键点检测是对野生动物进行姿态追踪的重要步骤。基于深度卷积网络(Deep Convolutional Neural Network,DCNN),首次使用Transformer模型实现野生动物的关键点检测任务,Transformer可以捕捉野生动物关键点的长距离依赖关系。在网络的基本模块(BasicBlock)中引入SGE(Spatial Group-wise Enhance)注意力机制,改善所提取特征的分布。提出了一种基于空间注意力的多分辨率表征融合方法,在特征融合时关注更有用的特征信息。在野生动物关键点检测的公共数据集上进行实验,实验结果显示,所提方法获得了良好的效果。

Endangered wildlife should be effectively protected.Pose tracking is a feasible way to protect wildlife.The wildlife pose estimation is an important step to pose tracking.Based on the deep convolutional neural network(DCNN),transformer is adopted in the wildlife pose estimation task for the first time.The transformer model can capture the long-range dependency between the key points of wildlife.SGE(spatial group wise enhancement)attention mechanism is introduced for better perfor-mance.A new-built basicblock is designed to improve the distribution of extracted features.A multiresolution representation fusion measure based on spatial attention focuses on more effective features during aggregating process.The method is verified on the dataset of wildlife pose estimation and experimental result is promising.