C 8.2 | Faster RCNN - Reducing Proposals from Region Proposal Network | Machine Learning | EvODN

Considering that, we might end up with around 20,000 proposals from RPN, the question is, how do I reduce the number of proposals while still maintaining the quality.
For this, we can pick a hint from Selective Search and Edge Boxes, which basically eliminate the Background regions of the image from the proposals.
But how to do that with the RPNs?
Simple, we add a classifier for every BBox Regressor, which will tell me whether the region is a foreground or background region. Unlike the classifiers in the earlier discussions, this need not differentiate between many classes.
And if we use softmax, we will also get the rating of the foreground regions. So, we can also rank the foreground regions based on the score. The advantage is that, now it allows us to pick the best few proposals. And experimentally, we can decide how many proposals would be good enough.
During training, we train the Classifier with just Foreground or Background labels.
So, this is how we can reduce the number of proposals.
Questions that are remaining is whether it is faster and more accurate than SS/EB. Else, all this design would be of no use. To make it faster, the RPN needs one more refinement. Which we will see in the next video.
------------------------
This is a part of the course 'Evolution of Object Detection Networks'.
See full playlist here: • Evolution Of Object De...
------------------------
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.

[@Cogneethi]

See full course on Object Detection: th-cam.com/play/PL1GQaVhO4f_jLxOokW7CS5kY_J1t1T17S.html
If you found this tutorial useful, please share with your friends(WhatsApp/iMessageMessenger/WeChat/Line/Telegram) and on Social(LinkedIn/Quora/Reddit),
Tag @cogneethi on twitter.com
Let me know your feedback @ cogneethi.com/contact

[@adityavaishampayan1975]

Dude, you rock

[@randalllionelkharkrang4047]

Im confused, whats the difference between bounding box , anchors and reference box?

[@dingusagar]

Are there 9 different background classifiers one for each of the 9 bbox regressors?

[@praneethcj6544]

Very well explained

[@shonischannel5405]

How is ground truth bounding box generated?

[@Cogneethi]

its manually done using some tools like labelme etc

[@shonischannel5405]

@@Cogneethi thank you for replying...can i get your mail id so that i can clear few doubts...