Transcript
Page 1: D tour baskar-rethinasabapathi_presenter

d.tour: Stylebased Exploration ofDesign Example Galleries

Baskar Rethina-saba-pathi

Page 2: D tour baskar-rethinasabapathi_presenter

Problem at hand

Beginning a design work Design conceptualization is usually abstract

Need example designs to adapt / refer / copy

Existing methods

Keyword-only web search Difficult to articulate exact keywords

Results based on content, not style

Search through template designs Difficult to measure the usability

Page 3: D tour baskar-rethinasabapathi_presenter

Prior work

Rough layout Sketch

Designers don’t have concrete ideas to start with

Cannot explore non-layout dimensions

Comprehensive design gallery

Limited only to browsing ? search, recommendation, filtering, exploration

Page 4: D tour baskar-rethinasabapathi_presenter

Answer is d-tour

https://www.youtube.com/watch?v=g4ggoccLOyI

Page 5: D tour baskar-rethinasabapathi_presenter

d-tour

What is it ?

an exploration tool for web designs

search by example design or keyword + various parameters

from a curated database of webpages

Key aspect

Matches the stylistic similarity

Page 6: D tour baskar-rethinasabapathi_presenter

Feature Selection

Which feature Based on discussions with designers and from design literature

How

Global features directlyfrom DOM tree

Bento page Segmentationfor local features

Page 7: D tour baskar-rethinasabapathi_presenter

Feature mapping

Bricolage Algorithm ( other paper)

Map DOM trees of source and target pages

Define ‘Stylistic Distance’ as cost

Build a ‘Design Vector’ of costs

Page 8: D tour baskar-rethinasabapathi_presenter

Alternative approach?

Why not treat entire webpage as one image?Vision based approach?

Example: Google Image Search

CBIR techniques

Page 9: D tour baskar-rethinasabapathi_presenter

Arguments

CBIR techniques more suited for natural images Scene recognition

Object recognition

Difficult to capture style and structure Needs manual annotation again

Page 10: D tour baskar-rethinasabapathi_presenter

Critique on Query Processing Traditional Vector space model

Assumes equal weights for all features (?) If I query “blue + boxed”, who gets more importance – color or structure?

Features treated orthogonal No check for interdependence “Clumsy” can mean scattered text as well as non- agreeable colors

Use of space + Use of color

Text/Keyword Search

TF-IDF model - Uses Document’s text content

Assumes query and document content to be same I say “grey”, but what I really mean is “neutral”

Page 11: D tour baskar-rethinasabapathi_presenter
Page 12: D tour baskar-rethinasabapathi_presenter

Activity

Describe the design aspects of the websiteIn terms of style tags . E.g – large font, colorful etc.

Exchange your style tags (not the website)

Find the site matching the style tags (or a similar looking site)

Page 13: D tour baskar-rethinasabapathi_presenter

Style Tags used in d-tour

Page 14: D tour baskar-rethinasabapathi_presenter

Evaluation of d-tour

40 Participants – 20 Univ Studs + 20 Mturk Workers

Task Give example websites for a ‘School Science Education Program’

Randomly assigned either to use Web search or d tour

Write Why you chose this?

Page 15: D tour baskar-rethinasabapathi_presenter

Critique on Evaluation method

d tour users Feedback was generally positive with a few suggestions

Shown a tutorial of how to use d-tour Probably used the same style tags to search

Results are from a collection of best examples (ilovetypography.com, sixrevisions.com, and the Alexa Top100 )

Style biasing relevance ?

Web search users restricted to topic based search

Can attach an image + keyword in Google Image Search Do thumbnails better describe the webpage?

No School Students reviewed the results

Page 16: D tour baskar-rethinasabapathi_presenter

Discussion

1. Building effective databaseCrowd inputs help, but need decent database to attract crowd

2. Taste for designs – User specific?A good design to me is not even a design to youTrusting Ranking methods

3. Machine based approachNeed some image processing/machine learning to refine the process?

4. Extensive DOM dependency


Recommended