C 8.7 | Faster RCNN Demo - Pixel Norm, Conv Feature Maps and RPN BBoxes | CNN | Machine Learning скачать в хорошем качестве

C 8.7 | Faster RCNN Demo - Pixel Norm, Conv Feature Maps and RPN BBoxes | CNN | Machine Learning

In this demo, you will see the initial preprocessing state, where you pixel normalize the image and scale the smallest side to 600. Next, I show some of the Feature Map outputs from the 512 Feature Maps of VGGNet architecture. Then, I show how the Anchor Boxes are generated for the image. You can see that the boxes exceed the dimension of the image and the Receptive field. But the RPN has no problem refining the proposal. These proposals are then clipped to the image boundary. Then we take the top 300 proposals from the 17100 proposals. To do this, we first take the top 6000 proposals, then apply NMS. After NMS, for this image, you are left with around 600 proposals. Then, we just pick the top 300 based on the Softmax scores. These proposals are forwarded to the Fast RCNN network for BBox refinement and classification. ------------------------ 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.