C4W3

Question 1

Difference between classification with localisation vs object detection?

Answer 1

In object detection we have multiple objects

Question 2

Explain YOLO algorithm?

Answer 2

YOLO stands for you only look once
1. Place grid onto your image
2. You should run your network once so it will output a 1x1xN vector for each grid cell

Question 3

What is non max suppression?

Answer 3

Used to suppress multiple detection of the same object

Suppress regions (outputs) with less probability

Question 4

When anchor boxes are used?

Answer 4

When you want to distinguish overlapping objects

Question 5

What is regions proposal?

Answer 5

Use some clustering or another NN to propose regions to conv net

Question 6

What is intersection over union?

Answer 6

Metric for evaluating correctness of the predicted region

Question 7

How intersection over union works?

Answer 7

1. Compute size of the intersection
2. Divide by the size of the union
3. The pred is correct if >=0.5

Question 8

What is semantic segmentation?

Answer 8

When you want to classify each pixel on the image

Question 9

Idea of semantic segmentation and U-Net (Unit-Net)

Answer 9

Use transpose conv, which blows up smaller inputs
Use symmetric skip connection from initial layers to the latest convoluted layers

Question 10

Steps of non max suppression?

Answer 10

1. Discard all the boxes with <threshold prob
2. Discard any boxes with IOU >0.5 with boxes from step 1