Seeing, the review.
Seeing in twelve concepts, from filters sliding over a photo to the points of a hand, and six questions that bring the lessons together. If an answer doesn't add up, under the explanation you'll find the chapter to read again.
The key concepts.
- Convolution
A small filter, for example 3 by 3 pixels, that slides over the whole image and lights up where it finds its pattern. A network made of layers of filters is called convolutional, or CNN.
- Confidence
The percentage the model gives to an answer. It tells you how much the model prefers that answer over the others it knows, not how likely it is to be right.
- Shortcut
When a model learns a detail that comes along with the answer instead of the answer itself, for example the snow instead of the wolf. On new data without that detail it gets things wrong.
- Imbalanced classes
When one class has many more examples than the others. A model can look accurate by always giving the biggest class, so you have to check the mistakes class by class.
- Nothing class
An extra class with examples of when the object isn't there, like the empty table or a hand passing by. It gives the model a right answer for those moments too.
- Data leakage
When nearly identical examples end up both among the ones for learning and among the ones for testing. The model recognizes them, and the score looks better than it will really be.
- Data augmentation
Changing the examples a little at random during training, for example by rotating a photo or making it brighter. The model sees more variety without you having to collect more data.
- IoU
Intersection over union. How much two boxes overlap, from 0 to 1. It's used to decide whether a box found by the model matches the right one.
- NMS
Non-maximum suppression. Among the boxes that overlap too much it keeps the one with the highest confidence and deletes the others, so one box is left for each object.
- Instance segmentation
A segmentation that separates each single object with its own mask, instead of coloring all the objects of the same class the same way.
- Depth estimation
A model that gives each pixel a distance instead of a class. From a single photo it estimates what's in front and what's behind, but not the exact meters.
- Normalization
Bringing the data to a common scale before giving it to the model. For the hand points you put the wrist at zero and divide by the size of the hand, so only the shapes are left.
6 questions that weren't in the lessons.
They bring the course's lessons together. After each answer there's the explanation and the place to read again. The review doesn't change the map.
-
MobileNet's body already knows how to see. You want to tell apart two olive varieties that look a lot alike. How many photos do you need?
Twenty photos are a starting point. Classes that look alike need many more. Train with 10, 20, 40 and 80 photos per class and see if the accuracy on the test photos is still going up. Explained in 02.02, Go deeper →
-
The model gets almost all the test photos right, then on a friend's table it often gets it wrong. Why didn't the test photos warn you?
If test and training photos come from the same session, the score gets inflated and doesn't tell you how it will do elsewhere. Split by session and take photos in different places. Explained in 02.02, chapter 6 →
-
Your gesture model says fist even when you scratch your nose. What do you add?
A model always chooses between the classes it knows. With a nothing class, a hand doing something else also has its own right answer, instead of a wrong fist. Explained in 02.02, chapter 4 →
-
Two apples close together end up in a single box. Which threshold do you change?
NMS keeps the best box and deletes the ones that overlap it too much. With a low IoU threshold it also deletes the box of the apple next to it. Raise it and both boxes stay. Explained in 02.03, chapter 6 →
-
You only need to know whether a leaf is healthy or sick. What kind of answer do you ask the model for?
The more precise the answer, the more it costs in calculations and in labels. A class takes one tap to assign. Boxes and masks are for when it matters where the spot is or how big it is. Explained in 02.04, chapter 7 →
-
You record all the fists in the bedroom and all the open hands in the kitchen. Could the gesture model learn the background?
The ready-made model turns each frame into 63 numbers. The background and the light stay behind in the pixels, so your model never sees them and can't use them as a shortcut. Explained in 02.05, chapter 4 →