论文阅读：Cut, Paste and Learn: Surprisingly Easy Synthesis for Instance Detection

1、论文总述

本篇论文是一篇对室内场景的目标检测数据集进行合成（Cut->Paste）的操作，通过这种合成操作进行数据增广，区别于一般意义上的数据增强操作，概括来讲就是：通过一个FCN网络对目标检测数据集里的bbox标注进行分割，学到mask，通过这个mask将前景和背景进行分离，然后将这个mask进行数据增广（旋转、遮挡、裁切、false positive），最后将这些mask贴到不同的背景上，由于贴的时候前景与背景的边界上有人工的痕迹（不那么自然），所以作者提出利用泊松或者高斯滤波对这些 边界进行处理，将不同方法处理过后的数据都扔进网络里进行训练，从而让网络忽略这个边界不自然的特征。
注： 看这篇论文是因为项目需要，因为最近的病害检测项目里也需要对数据进行合成，因为某个类别的数据太少了，人工去收集数据非常困难，所以暂时用这种人工合成的数据方法。

上图中提到的different blending 就是指处理paste时的边界时用不同的方法，然后都扔进网络里去训练，好让网络忽略这个特征。

Data generated using our approach is surprisingly effective at training detection models. Our results suggest that
curated instance recognition datasets suffer from poor coverage of the visual appearances of the objects. With our
method, we are able to generate many such images with
different viewpoints/scales, and get a good coverage of the
visual appearance of the object with minimal effort. Thus,
our performance gain is particularly noticeable when the
test scenes are different from the training scenes, and thus
the objects occur in different viewpoints/scales.
当测试场景与训练场景有区别时，本方法提升更多

2、 feature rich and feature poor objects

Instance detection is a well studied problem in computer vision. [55] provides a comprehensive overview of
the popular methods in this field. Early approaches, such
as [6], heavily depend on extracting local features such as
SIFT [30], SURF [3], MSER [32] and matching them to
retrieve instances [29, 48]. These approaches do not work
well for objects which are not ‘feature-rich’, where shapebased methods [10, 19, 21] are more successful.
Modern detection methods [14, 15, 39] based on learned
ConvNet features [23, 25, 44] generalize across feature
rich and feature poor objects [43].

有空看：参考文献[10] V. Ferrari, T. Tuytelaars, and L. Van Gool. Object detection
by contour segment networks. In European conference on
computer vision, pages 14–28. Springer, 2006. 2

3、目标检测与实例检测

4、Adding Objects to Images时的边界处理

问题：

After automatically extracting the object masks from input images, we paste them on real background images.
Na¨ıvely pasting objects on scenes results in artifacts which
the training algorithm focuses on, ignoring the object’s visual appearance

不同的边界处理方法：

消融实验：

参考文献：

Cut, Paste and Learn: Surprisingly Easy Synthesis for Instance Detection论文