基于混合边界框评分的实例分割模型

INSTANCE SEGMENTATION MODEL BASED ON HYBRID BOUNDARY BOX SCORING

YOLACT是实例分割中速度最快的算法之一,其分类置信度和定位准确度的低相关性会导致定位以及分割失败,针对这一问题,提出一种混合边界框评分的方法。在YOLACT的基础上,并行添加一个预测边界框交并比的分支,使用滑动平均绝对误差损失函数对此训练,将区域的交并比和分类置信度相乘作为边界框的评分,用此评分进行非极大值抑制,保留定位准确且分类置信度高的边界框;在特征金字塔网络的基础上添加一层自下而上的特征融合,增强了定位能力。在COCO2017和Pascal 2012 SBD测试集的实验结果表明,该方法将YOLACT的平均分割精度分别提高了3.2%和3.0%。

YOLACT is one of the fastest instance segmentation algorithms.However,the low correlation between the classification score and localization accuracy of the predicted detections will cause localization and segmentation failures.A hybrid bounding box scoring method is proposed to solve this problem.On the basis of YOLACT,a branch for predicting the intersection and union ratio of bounding boxes was added in parallel.The sliding average absolute error loss function was used to train this branch,and the intersection and union ratio of the region was multiplied by the classification confidence as the bounding box score.This score was used for non-maximum suppression,while retaining the bounding box with accurate localization and high classification confidence.Based on the feature pyramid network,a layer of bottom-up feature fusion was added to enhance the localization accuracy.The experimental results on COCO2017 and Pascal 2012 SBD dataset show that the proposed method can substantially improve AP of YOLACT by 3.2%on COCO test and 3%on PASCAL SBD 2012 test compared with the baseline.