Deep Learning For Dummies

Chapter 13

Learning Advanced CNNs

IN THIS CHAPTER

Understanding the importance of object detection

Distinguishing between detection, localization, and segmentation

Testing object detection by RetinaNet from a GitHub implementation

Realizing the weak spots of CNNs that could be exploited

Deep learning solutions for image recognition have become so impressive in their human-level performance that you see them used in developing or already marketed applications, such as self-driving cars and video-surveillance appliances. The video-surveillance appliances already perform tasks, such as automatic satellite image monitoring, facial detection, and people localization and counting. Yet you can’t imagine a complex application when your network labels an image with only a single prediction. Even a simple dog or cat detector may not prove useful when the photos you analyze contain multiple dogs and cats. The real world is messy and complex. You can’t expect, except in limited and controlled cases, laboratory style images that consist of single, clearly depicted objects.

The need to handle ...