Efficient Segmentation: Learning Downsampling Near Semantic Boundaries学习语义分割边界附近的下采样

Abstract:
为提高语义分割执行速度,会损失一些小物体,降低准确率。提出了新的自适应内容下采样技术(content-adaptive downsampling technique) 学习目标边界附近的采样位置。有利于提升分割边界质量和smaller-size object。
Introduction
Efficient Segmentation: Learning Downsampling Near Semantic Boundaries学习语义分割边界附近的下采样

non-uniform downsampl不均匀采样
提出的content-adaptive sampling的语义分割体系,分为3部分:1.non-uniform downsampling block训练目标类语义边界附近的采样像素。2.下采样分割模型:现存的分割模型。3.对分割结果上采样生成原始分辨率的分割图。

Efficient Segmentation: Learning Downsampling Near Semantic Boundaries学习语义分割边界附近的下采样
训练一个小型辅助网络预测sampling tensor without boundaries。辅助网络:两个Unet子网络堆叠,子网络堆叠动机是:模拟边界计算和采样点选择的连续过程(所以到底是???)
Efficient Segmentation: Learning Downsampling Near Semantic Boundaries学习语义分割边界附近的下采样
非均匀下采样模块的结构:
Efficient Segmentation: Learning Downsampling Near Semantic Boundaries学习语义分割边界附近的下采样
几个分割数据集:ApolloScapes [22], CityScapes [19], Synthia [45]
and Supervisely (person segmentation) [two labels: person and background] datasets.

问题:不知道自适应是如何实现的?

上一篇:论文阅读 (三):An empirical study on image bag generators for multi-instance learning (2016)


下一篇:140-CVPR论文阅读:《D2Det: Towards High Quality Object Detection and Instance Segmentation》