Day #186 - How Faster R-CNN, Fast R-CNN and R-CNN Works

Key Summary

• RPN Share CNN layers with detectors
• RPN can propose can scale / ratio
• RPN is scale invariant

R-CNN

• Input Image
• Identify Proposed Region (Through Selective Search)
• RCNN will run CNN for region proposals 
• Regional Proposal + CNN

Fast R-CNN

• CNN and ROI
• Estimate regional proposal by CNN
• Need external proposal

Faster R-CNN

• 250x faster than RCNN
• Doesn't need an external proposal
• Detector + Region Proposer

Architecture

• CNN - VGG-16 and ZFnet
• Faster RCNN uses set of parameters
• Kernel Filter / Sliding Windows
• Shared weights

Region Proposer

• CNN -> Feature map -> Proposed Regions
• IOU = Area of Intersection (Anchor and Ground Truth Box) / Area of Union  (Anchor and Ground Truth Box)
• Anchor - Mechanism to define area contains image or no using IOU
• IOU > 0.7 it contains object
• Faster RCNN will accept any size of image
• Sliding window to extract proposals using CNN
• 256 Positive Anchors
• 256 Negative Anchors
• From 2400 combinators
• Classfier to determine object / not
• Anchor / Ratio different type of anchors
• 128, 256, 512 scales
• Anchors are 21,600
• Get rid of low probability predictions
• Nail down on objects with classes (Non-Max Suppression)

\

RPN Loss Function

• Predicted probability, Find probability of object
• Consider positive anchors
• Minimize error between predicted and ground truth box

Fast RCNN - Selective external search
Faster RCNN

• Detector
• Has feature map
• Input image with RPN
• CNN train along with Regions Identified

Region Proposals

• Skip regions with no objects
• Regions with CNNs
• Pick few regions to run 
• Select few windows
• Segmentation Algo ran to figure out object (Blob)
• Take Blob Regions and Run CNNs
• Find blobs place bounding boxes on them 

Speed up Work

• Propose regions - Label + Bounding Box
• Convolution Implementation of Sliding windows to classify all proposed regions
• Clustering set to propose is Slow

Faster RCNN - CNN to propose regions instead of segmentation

Object Localization

• Where in the pic object is located
• Classification with Localization
• NN to output bounding boxes bx, by, bh, bw (parameterize bounding box)
• Label Training set with boundingbox
• BB use regression methods
• NN output bunch of real numbers
• Label landmark positions in training set, train them and output the NN

Happy Mastering DL!!!