Classic Meets Deep
Why learned features won.
For fifty years, vision engineers hand crafted features, and each one took years of research. Then, in twenty twelve, a network that learned its own features beat them all, by a huge margin. Some careers changed overnight. But the classic tools didn't die. They just found their proper place.
- Fifty years of hand-crafted features
- 2012: learned features won
- Classic tools found their place
From edges to objects
Here's why deep learning won. A convolutional network has the same filters that we've just learned about. But nobody designs them. The first layers learn edges and colours. Later layers combine them into textures and shapes. And the last layers respond to whole objects, like eyes and wheels. Everything is learned from examples, for the exact task at hand.
Classic or deep?
So what should you use? Use classic methods when lighting is controlled, the objects are simple, and you need speed, or very little data. Measuring, counting, and alignment are often easy this way. Use deep learning when appearance varies, as with real scenes, natural objects, or anything that's hard to describe with rules. And many of the best systems use both.
| Situation | Use |
|---|---|
| Fixed lighting, simple parts | Classic |
| Exact measurement | Classic + calibration |
| Natural, varied scenes | Deep learning |
| Tiny device, no GPU | Classic, or tiny model |
| Best overall | Often both |
Summary
Learned features beat hand crafted ones on hard problems, while classic tools remain perfect for simple ones. Our Deep Learning course builds CNNs from scratch. Here, let's use the very best ones, starting with a model that already knows a thousand objects.
Key takeaways- Networks learn their own filters
- Classic for simple, controlled tasks
- Often use both