用于人体实例分割的卷积神经网络

Convolutional Neural Network for Human Instance Segmentation

针对当前的实例分割算法无法分割两个高度重叠的人体对象,且量化的Mask实例与其ground truth之间的IoU的Mask质量通常与分类分数相关性不强等问题,利用人体骨骼和姿态来对人体进行分割,增加一个全新的Evaluation模块,利用预测Mask与ground truth之间的IoU来描述实例分割质量,提出了一种直接学习IoU的网络,能够提高实例分割的质量。为了获得更加丰富的特征信息,采用ResNet和FPN网络进行特征提取,融合多层特征的信息,使分割结果更加准确。实验结果表明,提出的网络框架对人体分割的结果更加准确,具有更加优越的鲁棒性。

Aiming at the current instance segmentation algorithm,it was impossible to divide two highly overlapped objects,the mask quality of the IoU wasn′t strongly correlated with the classification score.So human bones and poses are used to solve it.A brand-new Evaluation module is added to describe the quality of instance segmentation by the IoU between predicted Mask and ground truth.A network that directly learns IoU is proposed,which can improve the quality of instance segmentation.In order to obtain more abundant feature information,ResNet and FPN networks are utilized for feature extraction.The information of multiple features is merged to make the segmentation result more accurate.Experimental results show that the proposed network framework is more accurate and robust to human body segmentation.