本文转载自:
http://www.cnblogs.com/lillylin/p/6207292.html
SSD论文阅读(Wei Liu——【ECCV2016】SSD Single Shot MultiBox Detector)
目录
- 作者及相关链接
- 文章的选择原因
- 方法概括
- 方法细节
- 相关背景补充
- 实验结果
- 与相关文章的对比
- 总结
作者
- intro: ECCV 2016 Oral
- arxiv: http://arxiv.org/abs/1512.02325
- paper: http://www.cs.unc.edu/~wliu/papers/ssd.pdf
- slides: http://www.cs.unc.edu/%7Ewliu/papers/ssd_eccv2016_slide.pdf
- github: https://github.com/weiliu89/caffe/tree/ssd
- video: http://weibo.com/p/2304447a2326da963254c963c97fb05dd3a973
- github(MXNet): https://github.com/zhreshold/mxnet-ssd
- github: https://github.com/zhreshold/mxnet-ssd.cpp
- github(Keras): https://github.com/rykov8/ssd_keras
文章的选择原因
- 性能好,single stage
方法概括
-
文章的方法介绍
- SSD主要用来解决目标检测的问题(定位+分类),即输入一张待测图像,输出多个box的位置信息和类别信息
- 测试时,输入一张图像到SSD中,网络输出一个下图最右边的tensor(多维矩阵),对该矩阵进行非极大值抑制(NMS)就能得到每个目标的位置和label信息
Figure2的最右图的1th-20th Channel表示类别,每一个Channel上的map对应原图,last 4 channel的每一个map分别对应x,y,w,h的偏移量。最后4个通道可以确定一个box的位置信息,前20个通道确定类别信息。
-
方法的pipeline和关键点
方法细节
相关背景补充
实验结果
与相关文章的对比
总结
-
文章贡献
- SSD, a single-shot detector for multiple categories (faster than YOLO, accurate as Faster R-CNN)
- The core of SSD is predicting category scores and box offsets for a fixed set of default bounding boxes using small convolutional filters applied to multiple feature maps from different layers
- Experimental evidence: high accuracy, high speed, simple end-to-end training (single shot)
-
SSD对于其他方法的改进的关键点
-
- Using a small convolutional filter to predict object categories and offsets in bounding box locations
- Using separate predictors (filters) for different aspect ratio detections
- Using multiple layers for prediction at different scales (apply these filters to multiple feature maps to perform detection at multiple stages)