Perceptual Generative Adversarial Networks for Small Object Detection

2017CVPR 新鲜出炉的paper，这是针对small object detection的一篇文章，采用PGAN来提升small object detection任务的performance。

最近也没做object detection，只是别人推荐了这篇paper，看了摘要觉得通俗易懂就往下看了。。。最后发现还是没怎么搞懂，只是明白PGAN的模型。如果理解有误的地方，请指出。

言归正传，PGAN为什么对small object有效？具体是这样，small object 不好检测，而large object好检测，那PGAN就让generator 学习一个映射，把small object 的features 映射成 large object 的features，然后就好检测了。PGAN呢，主要就看它的generator。

传统GAN中的generator是学习从随机噪声到图像的映射，也就是generator可以把一个噪声变成图片，而PGAN的思想是让generator把small object 变成 large object，这样就有利于检测了。 来看看文章中的原话都是怎么介绍generator的：

1. we address the small object detection problem by developing a single architecture that internally lifts representations of small objects to “super-resolved” ones, achieving similar characteristics as large objects
2. Perceptual Generative Adversarial Network (Perceptual GAN) model that improves small object detection through narrowing representation difference of small objects from the large ones.
3. generator learns to transfer perceived poor representations of the small objects to super-resolved ones
4. The Perceptual GAN aims to enhance the representations of small objects to be similar to those of large object
5. the generator is a deep residual based feature generative model which transforms the original poor features of small objects to highly discriminative ones by introducing fine-grained details from lower-level layers, achieving “super-resolution” on the intermediate representations
   
   6.传统的generator G represents a generator that learns to map data z from the noise distribution pz(z) to the distribution pdata(x) over data x，而PGAN的generator中 x and z are the representations for large objects and small objects
6. The generator network aims to generate super-resolved representations for small objects to improve detection accurac
7. the generator as a deep residual learning network that augments the representations of small objects to super-resolved ones by introducing more fine-grained details absent from the small objects through residual learning

文章在不同地方不断的重复了一个意思，就是generator学习的是一个映射，这个映射就是把假（small object）的变成真(large object)的

来看看generator长什么样子

分两个部分，这里就没看懂是什么意思了，或许和object detection有关了。最终得出的结果是Super-Resolved Features 这个就很像Large Objects Featuresle. 如图，左下角是G生成的，左上角是真实的：

讲完了generator 就到discriminator了，这里的discrimintor和传统的GAN也有不一样的地方。

在这里，加入了一个新的loss，叫做perceptual loss ，PGAN也因此而得名（我猜的，很明显嘛）这个loss我也是没看明白的地方，贴原文大家看看吧（有理解的这部分的同学，请在评论区讲一讲，供大家学习）

1. justify the detection accuracy benefiting from the generated super-resolved features with a perceptual loss

看完paper感觉作者没有很直接说提出PGAN是inspired by哪些文章~不过GAN（2014 Goodfellow）