C 7.9 | Fast RCNN Network, Computation Time, Accuracy | CNN | Object Detection | Machine learning скачать в хорошем качестве

C 7.9 | Fast RCNN Network, Computation Time, Accuracy | CNN | Object Detection | Machine learning

Fast RCNN is an improvement on top of SPPNet. First change they did was to remove the multilevel pooling in the SPP layer. Instead they use a single 7x7 grid for SPP. Next, they realized that whether you use Softmax or SVM for classification, your accuracy is more or less same. So they got rid of the SVM classifier. Instead of separately training the Classifier first and then the BBox regressor as is usually done, they combined the losses from both and fine tuned the network upto the L3 conv layer. Why just L3? As we know, the initial layers of ConvNet extract some generic features and there would not be much to be gained by fine tuning these layers. Also, they made the FC6 and FC7 layers common to both Classifier and BBox Regressor. And they added an extra FC layer each to both of them. Lastly, they used Smooth L1 loss in BBox Regressor instead of L2 loss. If you want to know more about these loss functions, see: https://heartbeat.fritz.ai/5-regressi... With these changes, you get the Fast RCNN network. This network happens to be 143 times faster than RCNN. While SPPNet is around 20 times faster than RCNN. So, in terms of speed there has been a major improvement. However, on the accuracy front, there is not much change. In RCNN, the accuracy is 66%, here it is 66.9% ------------------------ This is a part of the course 'Evolution of Object Detection Networks'. See full playlist here:    • Evolution Of Object Detection Networks   ------------------------ Copyright Disclaimer: Under section 107 of the Copyright Act 1976, allowance is made for “fair use” for purposes such as criticism, comment, news reporting, teaching, scholarship, education and research.