Object Detection

Question 1

When there are multiple bounding boxes for an object in an image, what is Intersection over Union (IoU) ratio?

Answer 1

It is Intersection Area / Union area

Question 2

YOLO - In a NxN grid of an image, if each grid box predicts Output = [Pc, Bx, By, Bh, Bw, C1, C2, C3], what does each value stand for

Answer 2

Pc -> Probability the grid element contains an object
Bx, By -> midpoint of the object
Bh, Bw -> Height and Width of the object
C - Class of object

Question 3

If multiple grid elements output bounding boxes for the same object, how do we identify the most appropriate bounding box using Non-Max suppression?

Answer 3

• Discard all object with probability < 0.6
• Pick box with largest Pc
• Discard any other box with IoU >= 0.5

Question 4

What is object locaization?

Answer 4

It is determining the location and class of a single object in an image.

It is training and predicting [P,bx, by, bh, bz, c1,c2, c3] where P is probability of presence, b is coordinate, and c is the class of object.

Question 5

What is sliding windows for object detection?

Answer 5

Using sliding windows of multiple sizes and using each window in a CNN to classify it.

Question 6

Explain Yolo

Answer 6

Image is divided into 9 grid cells and object localisation is is applied in all 9 independently.

Question 7

In Yolo, what if an image is spread across grid cells?

Answer 7

We assign the object to grid where its midpoint falls