09. Faster R-CNN

08 Faster RCNN V1 RENDER V2

Region Proposal Network

You may be wondering: how exactly are the RoI's generated in the region proposal portion of the Faster R-CNN architecture?

The region proposal network (RPN) works in Faster R-CNN in a way that is similar to YOLO object detection, which you'll learn about in the next lesson. The RPN looks at the output of the last convolutional layer, a produced feature map, and takes a sliding window approach to possible-object detection. It slides a small (typically 3x3) window over the feature map, then for each window the RPN:

  1. Uses a set of defined anchor boxes, which are boxes of a defined aspect ratio (wide and short or tall and thin, for example) to generate multiple possible RoI's, each of these is considered a region proposal.
  2. For each proposal, this network produces a probability, Pc, that classifies the region as an object (or not) and a set of bounding box coordinates for that object.
  3. Regions with too low a probability of being an object, say Pc < 0.5, are discarded.

Training the Region Proposal Network

Since, in this case, there are no ground truth regions, how do you train the region proposal network?

The idea is, for any region, you can check to see if it overlaps with any of the ground truth objects. That is, for a region, if we classify that region as an object or not-object, which class will it fall into? For a region proposal that does cover some portion of an object, we should say that there is a high probability that this region has an object init and that region should be kept; if the likelihood of an object being in a region is too low, that region should be discarded.

I'd recommend this blog post if you'd like to learn more about region selection.

Speed Bottleneck

Now, for all of these networks including Faster R-CNN, we've aimed to improve the speed of our object detection models by reducing the time it takes to generate and decide on region proposals. You might be wondering: is there a way to get rid of this proposal step entirely? And in the next section we'll see a method that does not rely on region proposals to work!

Why is speed important?

Multi-object identification consists of locating and classifying different objects in an image. Some models trade off accuracy for speed. What kinds of applications require speed in object recognition? Check all that apply.
SOLUTION:
  • Pedestrian detection for an autonomous vehicles.
  • Tracking people's faces so that a camera can focus on them.

Faster R-CNN Implementation

If you'd like to look at an implementation of this network in code, you can find a peer-reviewed version, at this Github repo.

Next Concept