一种基于Interactive KeyPoints的目标检测算法

An Object Detection Algorithm Based on Interactive KeyPoints

目标检测是计算机视觉中的研究热点,小目标检测精度低一直是难点。论文采用关键点的预测方式和回归方法,建立一种基于交互关键点(Interactive KeyPoints)的无锚检测模型(ITKP)。该模型采用自适应采样方法获取一组表示对象的空间范围和具有重大语义意义的局部区域的关键点集后,通过自注意力(self-attention)机制建立关键点之间的联系以更好地匹配关键点和实现细粒度的定位。另外使用自注意力层代替卷积层来获取特征的全局上下文信息,设计了一个交互的特征金字塔,实现了更精确的识别和定位。论文提出的检测模型计算复杂度低,在基于关键点检测的无锚检测方法中,检测性能和小目标检测效果有较大的提升。同时保持实时的检测速度,为嵌入式或边缘设备的应用提供理论支持。

Object detection is a research hotspot in computer vision,and the low accuracy of small target detection has always been a difficult point.The keypoints prediction and regression method is adopted to establish an anchor-free detection model(IT⁃KP)based on Interactive KeyPoints.The model employs an adaptive sampling method to obtain a set of key point sets representing the spatial range of the object and the local area with significant semantic meaning,and establishes the connection between the key points through the self-attention mechanism to better match the keypoints and achieve fine-grained positioning.In addition,the self-attention layer is used instead of the convolutional layer to obtain the global context information of the features,and an interac⁃tive feature pyramid is designed to achieve more accurate recognition and location.The detection model proposed in the paper has low computational complexity.In the anchor-free detection method based on keypoints detection,the detection performance and small target detection effect have been greatly improved.At the same time,it maintains real-time detection speed and provides theo⁃retical support for embedded or edge device applications.