For nearly one or two years, CNN has developed rapidly in the detection of this piece, and the following details review the development of the entire CNN testing domain model, as well as the development of time performance.
First, RCNN process:
Extract region (off model) + Extract features (on model) + classifyregions according feature (SVM or Softmax)
Performance:
Precision:
Second, spp-net process:
Do conv First, then extract features according to window. Why can't rcnn do the same? The reason is that the max pool processing of SPP can better satisfy the feature expression of different scale window.
Performance:
The core idea is to do only one conv in the whole picture, which is consistent with the Overfeat
Precision:
Third, FAST-RCNN process:
The ROI layer pooling is introduced, as well as the Multi-task training classification and detection frame.
Performance:
Compared to Sppnet, Fast R-CNN trains Vgg163xfaster, tests 10xfaster, and are more accurate.
In addition, the idea of FC layer SVD is also proposed.
Vgg Time Performance analysis
Precision:
The improvement of fast r-cnn over sppnetillustrates that even though Fast R-CNN uses Single-scale training and TESTING,FI Ne-tuning the conv layers provides a large improvement in MAP (from 63.1% to66.9%). Traditional R-CNN achieves a mAP of 66%. These results arepragmatically valuable given how much faster and easier Fast r-cnn are to Trainand test, which we discuss Next.
Iv. FASTER-RCNN Process:
On the basis of fast-rcnn, borrowed from the fcn of ideas, the proposal stage into a layer added to the network to learn.
Performance:
Cost-free for Proposal
Precision:
Our detection system have a frame rate of5fps (including all steps) on a GPUs, while achieving State-of-the-art Objectdetect Ion accuracy on PASCAL VOC (73.2% map) and (70.4% map) using300 proposals per image
Copyright NOTICE: This article for Bo Master original article, without Bo Master permission not reproduced.
CNN Test Summary