Hacksession Image Recognition

Dr. Thorben Jensen

Outline

07.11.2019

1 Intro Image Recognition

2 Object Detection and YOLO

3 Session Targets

4 Setup & Code Intro

2

Intro Image Recognition

Image classification in a nutshell

4Image source: https://www.mathworks.com

Images are matrices

5

Filters can match patterns

6

Applying multiple filters

7

To detect more complex features: apply filters after filters

8Image source: https://www.slideshare.net

Low-level Mid-level High-level Result

T. Cruise: 99 %

Input

T. Jensen: 1 %

Object detection & YOLO

Types of Image Recognition Tasks

07.11.2019 10

YOLO – „you only look once“

Conventional Object Detection

separately proposes and classifies ‘boxes’

YOLO (“you only look once”)

parallelizes proposing and classifying ‘boxes’

07.11.2019 11

https://towardsdatascience.com/r-cnn-fast-r-cnn-faster-r-cnn-yolo-object-detection-algorithms-36d53571365e

Network pre-trained on COCO-Dataset

• 80 object classes

• 330.000 images

• Networks pre-trained on COCO dataset freely available

07.11.2019INFORMATIONSFABRIK – Businesspräsentation 12

Session Targets

07.11.2019 15

https://56f2a99952126.streamlock.net/833/default.stream/playlist.m3u8

1. Form groups (2-3 persons each, 1 Google account per group)

2. Intro into sample code

3. Diving into sample code with Google Colab

4. Choose a new use case, and code it with your group

– How many bicycles?

– Delivery trucks? (trucks only allowed at limited hours)

– Remove certain objects from an image?

– < your idea here >

5. 16:30-16:45: present your results to us

07.11.2019 16

Agenda

Setup & Code Intro

Contact

07.11.2019 18

Dr. Thorben Jensen Data Scientist

+49 160 69 666 42

tjensen@informationsfabrik.de

www.informationsfabrik.de