R-CNN
[] R-CNN deep learning and personal papers reading comprehension
[] RCNN target detection algorithm Detailed
R-CNN paper Translation Studies
In short, R- CNN success of two key factors:
1 : bottom-up use of large convolution neural network (CNNs) on the candidate area for positioning and segmentation of the object.
2: When the training data is tagged, first tune supervised pre-training, and then to a particular task for auxiliary tasks, can produce significant performance gains
Briefly, RCNN the following four steps to achieve target detection:
. A determination about 1000-2000 in the image frame candidates
b for each candidate frame image blocks, used. Deep Web extracting feature
c of the extracted candidate box. features using a classifier determines whether a particular class
d. for the characteristics belonging to a candidate block with regressor further adjust its position
The classic R- several challenges CNN: Training division multiple steps (first pre-trained in the classification of the data set, then Fine - Tune training and then training for each category a linear SVM classifier, the last and then regressors regression of the bounding box, and the bounding box needed generated by a selective search) Time and space overhead big (in training SVM and return characteristics of the network when the need to use training as input features stored on disk read time into the larger overhead) Test slower (each image of each region proposal have to do convolution operation is repeated too many) Fast RCNN been made before SPPnet to solve R -CNN repeated convolution problem, but there are still SPPnet and RCNN similar defects: Training division multiple steps (SVM classifiers require additional regressors) Large space overhead
FAST-CNN
Target detection algorithm Fast R-CNN Introduction
Improvement
- Convolution is not repeated for each region proposal, but for the whole image before extracting feature generalization, this way reducing the amount of computation (note, R-CNN in for each region proposal will convolve there are a lot of double counting)
- ROIPooling proposed ingenious solution to the problem of scale zoom
- The regressor into the network training together, but instead of SVM classifier with softmax, more simple and efficient
Faster-rcnn
faster-RCNN key points (recommended area network RPN) Comments