Computer Vision Problems презентация

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

Computer Vision Problem
vertical edges
horizontal edges

Vertical edge detection
1
1
1
-1
-1
-1
0
0
0
1
1
1
-1
-1
-1
0
0
0

Vertical edge detection examples

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

Summary of convolutions
padding p
stride s

Multiple filters
6 x 6 x 3
4 x 4
3 x 3 x 3
3 x 3 x 3
4 x 4

Pooling layer: Max pooling

Pooling layer: Average pooling

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

Outline
Classic networks:
LeNet-5
ResNet
Inception
AlexNet
VGG

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

AlexNet
=
9216
4096
4096
MAX-POOL
MAX-POOL
3 3
MAX-POOL
Softmax
1000
[Krizhevsky et al., 2012. ImageNet classification with deep convolutional neural networks]

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

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

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

Classification with localization
1 - pedestrian
2 - car
3 - motorcycle
4 - background

Defining the target label y
1 - pedestrian
2 - car
3 - motorcycle
4 - background
Need to output class label (1-4)

Sliding windows detection

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

Non-max suppression example

Non-max suppression algorithm
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
Each output prediction is:

Non-max suppression example
0.8
0.7
0.6
0.9
0.7