Computer Aided Design for Advertisment Compositions

The growing e-commerce market is closely associated with an increasing demand for online advertisements. Graphical designers are scarce and Creative agencies are looking for solutions to keep up with the demand for advertisements by customers. Data Science has been applied successfully in the marketing domain already, but rarely on the creation of online advertisements.

Kay Hoogland will speak about the research steps to find a scalable data science solution for this problem. The research resulted in a tool that co-creates product compositions for online advertisements with designers. The intelligence of the tool is written in DEAP and runs serverless on AWS using Zappa.

Transcript

1. Computer-aided design of advertisement product compositions 15.05.2018 Kay Hoogland
2. Outline • Introduction • Problem • Mistakes • Approach • Results
3. INTRODUCTION
4. About me • Bachelor – Industrial Engineering – TU/e • Big Data Honors program (2 years) • Master – Data Science & Entrepreneurship – JADS • Master thesis @ Greenhouse Group
5. About Greenhouse Group • Eindhoven – November 2006 • Digital marketing umbrella company • Blue Mango Interactive, Fresh Fruit Digital, Pubnxt, We Are Blossom, and LemonPi • Multiple hubs • Technology, data, creative, media, research
6. About Greenhouse Group • Greenhouse Group Labs • Entirely consists of students • Research to potential disruptive technologies
7. PROBLEM
8. Better AI Cloud computing Growing ad market Online growth • Demand for design of banners Customer demand for relevant ads Repetitive work • Need for scalable job automation
10. Model
11. Design process 1. Read retailer’s requirements 2. Copy previous template 3. Sort out the images 4. Load in Photoshop 5. Scale images correctly 6. Compose images 7. Put pancake on the corner of composition 8. Check if composition ‘works’ 9. Apply reflection & drop shadow
12. What?
13. Why me?
14. MISTAKES
15. Data • ~200 compositions available • Adobe Photoshop files • Exported image files
16. Adobe Photoshop data Goal: • Extract the composition • Learn design rules
17. Adobe Photoshop data • Big files • Lack of proper naming conventions • No proper tooling
18. Adobe Photoshop data INFEASIBLE
19. Image data Goal: • Extract image features
20. Image data VGG19 Convolutional Neural Network • Weights trained on ImageNet Image features!
21. Features plotted using t-SNE
22. Features plotted using t-SNE
23. Image data INFEASIBLE
24. APPROACH
25. Interactive furniture layout Merrell, Schkufza, Li, Agrawala, Vladlen Koltun, 2011)
26. COMPUTER-AIDED DESIGN
27. Genetic algorithms
28. Initial experiment
29. Outcome
30. Research Design
31. Tools used Python: • DEAP – Genetic Algorithms • PILLOW – Create the compositions Web interface: • AWS EC2 instance (not Zappa) • Flask • JavaScript
32. Pasting images x y
33. For 4 items, assign a random (100-1200) x-coordinate and a random (100-500) y-coordinate = , , , , , , , Repeat until the population is filled (100). Population
34. Evaluation • Overlap • Whitespace • Bottom / side • Floaters • Smaller items in front
35. Selection (tournament) Select k individuals at random from the population, pick the one with the best fitness. Repeat until population is filled. k Winner
36. , , , , , , , , , , , , , , Parents , , , , , , , , , , , , , , Children Crossover (two-point)
37. Mutation Based on: • Component similarity • Height difference
38. Mutation • Add noise to coordinates of item
39. Mutation • Add noise to coordinates of item • Align two items vertically
40. Mutation • Add noise to coordinates of item • Align two items vertically • Align two items horizontally
41. Mutation • Add noise to coordinates of item • Align two items vertically • Align two items horizontally • Mirror two items
42. Mutation • Add noise to coordinates of item • Align two items vertically • Align two items horizontally • Mirror two items • Connect two items
43. Web Application
44. Designer feedback • Smaller items to the front • Align the composition in the middle • Group similar objects together • Variation between compositions • Depth difference
45. Balancing Convergence Variation
46. RESULTS
47. Designer feedback • 3 versions of the application tested • Satisfaction measured with 4 seven-point Likert items
48. Outcome 1 4 2 5 3 6
49. Pairwise evaluation • Set up this week
50. Thesis @ Greenhouse Group
51. Project • Video is going to play a huge role in marketing • Greater retain rate in video compared to static images • Problem: video is not scalable • Assignment: AI generated video | programmatic product / brand placement
52. SUMMARY
53. Summary • Mistakes • Adobe Photoshop data • Image data • Approach • Actively involve designers in the process • Improve the product iteratively • Results