Computer Vision Problems

1.

Computer Vision Problems
Image Classification
Neural Style Transfer
Cat? (0/1)
64x64
Object detection
Andrew Ng

2.

Deep Learning on large images
Cat? (0/1)
64x64
Andrew Ng

3.

Computer Vision Problem
vertical edges
horizontal edges
Andrew Ng

4.

Vertical edge detection
3
1
2
0
4
2
1
1
1
01
0
5
7
1
2
4
01
01
0-1
1
8
2
3
1
5
0-1
0-1
2
9
5
1
6
2
-1
-1
-1
7
3
1
7
2
3
4
1
3
8
8
9
Andrew Ng

5.

Vertical edge detection examples
10
10
10
0
0
0
10
10
10
0
0
0
10
10
10
0
0
0
10
10
10
0
0
0
10
10
10
0
0
0
10
10
10
0
0
0
0
0
0
10
10
10
0
0
0
10
10
10
0
0
0
10
10
0
0
0
10
0
0
0
0
0
0
0
30
30
0
1
0
-1
0
30
30
0
1
0
-1
0
30
30
0
1
0
-1
0
30
30
0
0
-30
-30
0
10
1
0
-1
0
-30
-30
0
10
10
1
0
-1
0
-30
-30
0
10
10
10
10
10
10
1
0
-1
0
-30
-30
0
Andrew Ng

6.

Valid and Same convolutions
“Valid”:
“Same”: Pad so that output size is the same
as the input size.
Andrew Ng

7.

Summary of convolutions
padding p
stride s
Andrew Ng

8.

Multiple filters
3x3x3
4x4
6x6x3
3x3x3
4x4
Andrew Ng

9.

Pooling layer: Max pooling
1
3
2
1
2
9
1
1
1
3
2
3
5
6
1
2
Andrew Ng

10.

Pooling layer: Average pooling
1
3
2
1
2
9
1
1
1
4
2
3
5
6
1
2
Andrew Ng

11.

Types of layer in a convolutional network:
- Convolution
- Pooling
- Fully connected
Andrew Ng

12.

Outline
Classic networks:
• LeNet-5
• AlexNet
• VGG
ResNet
Inception
Andrew Ng

13.

LeNet - 5
7
avg pool
avg pool
f=2
s=2
f=2
s=2
120
[LeCun et al., 1998. Gradient-based learning applied to document recognition]
84
Andrew Ng

14.

AlexNet
MAX-POOL
MAX-POOL
MAX-POOL
33
=
9216
[Krizhevsky et al., 2012. ImageNet classification with deep convolutional neural networks]
4096
4096
Softmax
1000
Andrew Ng

15.

VGG - 16
CONV = 33 filter, s = 1, same
POOL
[CONV 128]
2
POOL
224x224x 3
POOL
POOL
POOL
FC
4096
FC
4096
[Simonyan & Zisserman 2015. Very deep convolutional networks for large-scale image recognition]
Softmax
1000
Andrew Ng

16.

Inception network
[Szegedy et al., 2014, Going Deeper with Convolutions]
Andrew Ng

17.

What are localization and detection?
Image classification
Classification with
localization
Detection
Andrew Ng

18.

Classification with localization
1234-
pedestrian
car
motorcycle
background
Andrew Ng

19.

Defining the target label y
1234-
pedestrian
car
motorcycle
background
Need to output class label (1-4)
Andrew Ng

20.

Sliding windows detection
Andrew Ng

21.

Evaluating object localization
“Correct” if IoU 0.5
More generally, IoU is a measure of the overlap between two bounding boxes.
Andrew Ng

22.

Non-max suppression example
Andrew Ng

23.

Non-max suppression algorithm
Each output prediction is:
Discard all boxes with
While there are any remaining boxes:
• Pick the box with the largest
Output that as a prediction.
• Discard any remaining box with
IoU with the box output
in the previous step
Andrew Ng

24.

Non-max suppression example
0.6
0.8
0.9
0.7
0.7
Andrew Ng

25.

Anchor box example
Anchor box 1:
y =
Anchor box 2:
Andrew Ng