结合区域采样和类间损失的人体解析模型

Human Parsing Model Combined with Regional Sampling and Inter-class Loss

人体解析是一项细粒度级别的语义分割任务,随着人体解析数据集中标注类别的精细化,人体解析数据集呈长尾分布,导致对相似类别的识别难度不断增大。均衡采样是解决长尾分布问题的有效方法。针对人体解析任务中难以对标注目标进行均衡采样和模型对相似类别的误判率增加等问题,文中提出了一种结合区域采样和类间损失的人体解析模型,该模型包含语义分割网络、区域均衡采样模块(Regionally Balanced Sampling Module,RBSM)和类间损失模块(Inter-class Loss Module,ILM)3个部分。首先将待解析图片送入语义分割网络得到初步预测结果,RBSM对初步的预测结果和真实标签进行采样,对采样后的预测结果和真实标签计算主损失;同时提取出语义分割网络的最后一层特征热图与真实标签,并将其送入ILM计算类间损失,让模型同时优化主损失和类间损失,最终得到精度更高的模型。在MHPv2.0数据集上的实验结果表明,该模型在不更改原有语义分割网络结构的基础上将mIoU评测指标提高了1.3%以上,有效缓解了长尾分布和类间的相似性给人体解析带来的影响。

Human parsing is a fine-grained level semantic segmentation task.The refinement of annotated categories in the human parsing dataset makes the dataset follow a long-tailed distribution and improves the difficulty of identifying similar categories.Balanced sampling is an efficient way to solve long-tailed distribution problem,but it’s difficult to achieve balanced sampling of the labeled object in human parsing.On the other hand,the fine-grained annotation will make the model misjudge similar categories.In response to these problems,a human parsing model combined with regional sampling and inter-class loss is proposed.The model consists of the semantic segmentation network,regionally balanced sampling module(RBSM),and inter-class loss module(ILM).Firstly,the images are parsed by the semantic segmentation network.Next,the parsing results and the ground truth labels are sampled by regionally balanced sampling module.Then the sampled parsing results and sampled ground truth labels are utilized to calculate the master loss.Meanwhile,the inter-class loss between the heatmap features coming from the semantic segmentation network and ground truth labels are calculated in the inter-class loss module,and the master loss and the inter-class loss are optimized at the same time to get a more accurate model.Experimental results based on the MHPv2.0 dataset show that the mIoU of the proposed model improves by more than 1.3%without changing the structure of the semantic segmentation network.The algorithm effectively reduces the impact of the long tail distribution problem and similarity among categories.