一两年cnn在检测这块的发展突飞猛进,下面详细review下整个cnn检测领域模型的发展,以及在时间性能上的发展。
Extract region(off model) + extract features(on model) + classifyregions according feature (svm or softmax)
先做conv,再根据window提取特征。为什么rcnn不能也这么做呢?原因在于spp对不同尺度进行了max pool处理能更好的满足不同尺度window的特征表达。
核心思想在全图只做一次conv,这个和overfeat的思想一致
引入了ROI层pooling,以及multi-task同时训练分类和检测框。
Compared to SPPnet, Fast R-CNN trains VGG163× faster, tests 10× faster, and is more accurate.
另外还额外提出了fc层SVD的思想
Vgg时间性能分析
The improvement of Fast R-CNN over SPPnetillustrates that even though Fast R-CNN uses single-scale training and testing,fine-tuning the conv layers provides a large improvement in mAP (from 63.1% to66.9%). Traditional R-CNN achieves a mAP of 66.0%. These results arepragmatically valuable given how much faster and easier Fast R-CNN is to trainand test, which we discuss next.
在fast-rcnn的基础上,借鉴了FCN的思路,将proposal阶段转化成一个layer加进了网络一起学习。
cost-free for proposal
our detection system has a frame rate of5fps (including all steps) on a GPU, while achieving state-of-the-art objectdetection accuracy on PASCAL VOC 2007 (73.2% mAP) and 2012 (70.4% mAP) using300 proposals per image