FastMask: Segment Multi-scale Object Candidates in One Shot
CVPR2017
https://github.com/voidrank/FastMask
本文针对检测和分割问题提出 FastMask 实现 segment multi-scale objects in one shot
这里的 one shot （ original image）相对 Multi-shot（ image pyramid），Multi-shot 的缺点很明显，就是计算量很大。
候选区域提取包括 矩形框和分割两个类别方法
Bbox-based object proposal 都有哪些方法了? EdgeBox [31] and Bing [4],DeepBox [17],MultiBox [7],RPN
Segment-based object proposal: SelectiveSearch [25], MCG [1] and Geodesic [16],DeepMask[20], SharpMask [21]
Bbox-based proposal 和 Segment-based proposal相比较， scale 对 Segment-based proposal 的影响更大，a highly matched receptive field is demanded to distinguish the foreground object from background
4.1. Network Architecture
FastMask architecture
4.2. Residual Neck
对于怎么生成这个 特征图金字塔，我们分析了 Max pooling neck，Average pooling neck，Feed-forward neck，感觉效果都不好，最后提出了 Residual neck
4.3. Attentional Head
我们采用文献【20,21】的 head module for decoding mask and object confidence 效果不好，原因可能是 our feature pyramid is sparser in scales，我们提出了 Attentional Head
COCO validation set for box and segmentation proposals
尺度和性能关系
速度对比