Traffic sign recognition – 03 – algorithms in object detection (2)

Fast R-CNN (2015)

Base on successful of R-CNN, Ross Girshick suppose an extend to solve weakness problems of R-CNN, with a short title: Fast R-CNN. Weakness problems of R-CNN:

• Pipeline training contains multi steps:
• A large number of bounding box make high cost, also time to training too slow
• Low detect object

Strong point of Fast R-CNN is using only one single model instead of pipeline to detect region and classification at the same time.

From a picture, it exports a group of region proposals then put them into a deep CNN network. For example VGG-16 using to export features. The output of deep CNN network is a custom layer (Region of Interest Pooling – RoI Pooling), export features for a part of input picture.

After that, features will be connected to one layer. Finally, the output of model will contains 2 part. One for predict label throught a softmax layer and one for predict bounding box. This process will be run over and over again for each part of picture.

This model faster than R-CNN in both training and predict. But it still need a group of region proposal along with each input image.

You need to login in order to like this post: click here