Upload
darshan-santani
View
81
Download
0
Embed Size (px)
Citation preview
Loud and Trendy: Crowdsourcing Impressions of Social Ambiance in Popular Indoor Urban Places
Darshan Santani, Daniel Gatica-PerezIdiap and EPFL Switzerland
28 October 2015
ACM MM 2015, Brisbane
Which place looks more “romantic”?
A B
Which place feels more “loud”?
A B
Understanding Social Ambiance of Urban Places
Place impressions define our favorite hangouts and shape our new discoveries
Social ambiance elicited by perceptual cues and prior knowledge
Connections between psychological features of cities and key indicators like well-being and prosperity
Our goal: to provide “a better idea of how people perceive and experience places”
Brave New Idea?
Social Media
Mobile Technologies Lots of images
Lighting
Spatial Layout
Flooring
Wall Decorations
Table Layout
Ceiling
Holistic study of place ambiance would involve interpretation of multiple sources of information, including subtle visual (and audio) cues → Many
challenges for multimedia research
Open Issue: Multimedia approaches to study social perception of urban places.
Research Questions
● RQ1: What types of social media images are perceived as being more informative of the ambiance of popular indoor places?
● RQ2: Can the ambiance of an indoor place be reliably assessed by observers of social media images? If so, for what dimensions of ambiance?
Dataset – Places
Singapore
NYCSeattle
Mexico City
Barcelona Paris
50 popular places per cityPlaces: cafes, restaurants, bars, or clubs
Dataset – Places
Singapore
NYC
Seattle
Mexico City
BarcelonaParis
50 popular places per cityPlaces: cafes, restaurants, bars, or clubs
Dataset – Places
Singapore
NYC
Seattle
Mexico City
BarcelonaParis
50 popular places per cityPlaces: cafes, restaurants, bars, or clubs
Image Dataset #1
● Random Image Corpus● 50,000+ images for all 300 places via 4SQ API● 55% iPhone; 19% Android; 22% via Instagram● Randomly selected 3 images per place for a total of 900 images
Image Dataset #2
● Physical Environment Image Corpus● Clear views of the environment (manually chosen)● 3 images per place for a total of 900 images
RQ1: What types of social media images best convey the ambiance of popular indoor
places?
Use image corpora to judge which image selection approach results in images being perceived as more
adequate to convey ambiance.
Methodology
● Crowdsourcing via Mechanical Turk● 5 images – 2 from the Physical Environment image
corpus, and 3 from Random image corpus● Tasks:
● Rank the images based on how informative they are of the ambiance of a place
● Categorize the images in one of the four classes: – Food/Drinks, – People/Group, – Physical Environment, and – None of these.
Results
Images from Phy.Env. corpus are in Top 2 ranks
91.7% for ambiance
Results
Images from Phy.Env. corpus are in Top 2 ranks
91.7% for ambiance
96% of Phy. Env. Image corpus describe the physical environment
67% Random image corpus describe either food or people
Results
Phy. Env. image corpus describe the physical environment in 96.2% of the cases
Random image corpus describe food items or people in 67% of the cases
Finding from RQ1: Images with clear views of the environment are perceived as more suitable to characterize indoor
ambiance.
RQ2: Can the ambiance of an indoor place be reliably assessed by observers of social
media images? If so, for what dimensions of ambiance?
Methodology
● Crowdsourcing via Mechanical Turk
● Data: Phy. Env. Corpus 900 images across 300 places
● Ambiance Labels: 13 dimensions [Graham, 2011]
● Task: Rate personal impressions of the place ambiance along 13 dimensions on a 5-point Likert scale
● 10 annotations for each dimension per place → 3,000 responses.
[Graham 2011] L. T. Graham, & S. D. Gosling, “Can the ambiance of a place be determined by the user profiles of the people who visit it?”. In the Fifth International AAAI Conference on Web and Social Media (ICWSM), 2011
Artsy
Bohemian
Conservative
Creepy
Dingy
Formal
Sophisticated
Loud
Old Fashioned
Off the beaten path
Romantic
Trendy
Upscale
Ambiance Labels
Inter-annotator Consensus
[0.8, 1.0)
[0.6, 0.8)
[0.0, 0.6)
ICCs across Cities
[0.8, 1.0) [0.6, 0.8) [0.0, 0.6)Color Legend
Ambiance across Cities – Correlation Analysis
Comparison between Cities
● Relatively few significant differences across cities
● Popular places in Seattle are perceived as less artsy compared to places in Barcelona.
● Popular places in Paris are perceived as less old fashioned com-pared to NYC and Seattle.
Comparison between Cities
● Popular places in Seattle are perceived as less artsy compared to places in Barcelona.
● Popular places in Paris are perceived as less old fashioned com-pared to NYC and Seattle.
Findings from RQ2:
Reliable estimates of ambiance can be obtained using social media images, suggesting the
presence of visual cues to form place impressions.
Most aggregate impressions of ambiance are similar across popular places in all cities
Future Work
● What specific cues are used to form place impressions?● Color? Lighting? Spatial layout? Interior Design?
● Automatic recognition of ambiance● Establishing links between visual cues and place
ambiance
Q & A
Email: [email protected]: @SabMayaHai