# Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks

###### tags: `Paper reading`

## Paper Link
#### [Click here](https://arxiv.org/pdf/1506.01497.pdf)
---
## Abstract & Introduction

### Compare with others
### Introduce a Region Proposal Network (RPN)
> Fast R-CNN已經減少了檢測網絡的運行時間。然而proposals的計算仍是一個重要的瓶頸。
* shares full-image convolutional features with the detection network
> 因此region proposal基本不占用運算資源
* simultaneously predicts object bounds and objectness scores at each position.
> trained end-to-end
* merge RPN and Fast R-CNN into a single network
* by sharing their convolutional features—using “attention” mechanisms
* 引入了"anchor" box概念用於可以預測尺度和長寬比變化很大的regionproposal
> 

### Experiment Result
* frame rate of 5fps (including all steps) on a GPU
* achieving state-of-the-art object detection accuracy on PASCAL VOC 2007, 2012, and MS COCO datasets
> with only 300 proposals per image
---
## Method
### Architecture
### Region Proposal Networks

* input: image (of any size)
* outputs: a set of object proposals (with an objectness score)

#### Anchors
* reg 4k bbox的x,y,w,h
* cls 2k 每個proposal是目標或不是目標的概率
* 位於sliding window的中心，使用3個比例(0.5,1,2)和3個長寬比( 1:1,1:2,2:1)。
> 關於anchors size，其實是根據設定的base size(=16)設置的。
>

## Loss Function

---
## Refference
* [Faster R-CNN代码解析](https://zhuanlan.zhihu.com/p/61221686)
* [图解论文Faster RCNN](https://www.bilibili.com/video/BV13W411K7jM?p=6)