Download presentation
Presentation is loading. Please wait.
Published byBeverly McDaniel Modified over 8 years ago
1
City Forensics: Using Visual Elements to Predict Non-Visual City Attributes Sean M. Arietta, Alexei A. Efros, Ravi Ramamoorthi, Maneesh Agrawala Presented by: Manu Agarwal
2
Outline Introduction Related Work Methodology proposed in the paper Results Applications Discussion
3
One of these is from Paris This is Paris Clap if… Slide credits: Doersch, Singh
4
Clap if… Slide credits: Doersch, Singh
6
Modeling predictive relationships Higher greenery City appearance Cases of depression Higher housing prices Broken Windows Theory! Broken glass, graffiti, trashIncreased crime rate
7
Violent crime rate High housing prices
8
Related Work Doersch et al. – Method for identifying visual elements of a city that differentiate it from other cities – Binary classification Koller et al., Srinivasan et al. – Use video to track crowds for detecting flow patterns – Availability of such videos is limited
9
Problem Input: a set of measured (location, attribute- value) pairs and a set of (location, street-side panorama) pairs Output: Predictor that can estimate the value of a non-visual city attribute based on visual appearance
10
Methodology Spatially interpolate the input (location, attribute-value) data Build a bank of SVMs Build an attribute predictor from the resulting bank of SVMs using SVR
11
Interpolating Non-Visual City Attribute Values Use the radial basis function (RBF) r is the Euclidean distance between locations Authors use ɛ=2
12
Methodology Spatially interpolate the input (location, attribute-value) data Build a bank of SVMs Build an attribute predictor from the resulting bank of SVMs using SVR
13
Constructing the Visual Element Detectors 100k200k300k400k500k600k700k800k900k1M1.1M PositiveNegative Negative set Positive set
14
Constructing the Visual Element Detectors
15
Extract the set of image features in each panorama projection HOG+color features
16
Constructing the Visual Element Detectors Compute 100 nearest neighbors of features sampled randomly from the positive set PatchMatches
17
Constructing the Visual Element Detectors Compute an SVM for each nearest neighbor set Keep the top 100 SVMs
18
Methodology Spatially interpolate the input (location, attribute-value) data Build a bank of SVMs Build an attribute predictor from the resulting bank of SVMs using SVR
19
Computing the Predictor Features SVM scores
20
Computing the Predictor Retain the top 3 detection scores for each of the 100 SVMs
21
Computing the Predictor Estimate parameters w and b such that the following loss is minimized ɛ controls the magnitude of error that we can tolerate
22
Results: Analysis of Prediction Accuracy
24
Results: Prediction Maps
26
Results: Deep Convolutional Neural Network Features
27
Applications: Defining Visual Boundaries of Neighborhoods
29
Applications: Attribute-Sensitive Wayfinding
30
Applications: Validating Visual Elements for Prediction
32
Discussion Presence of other dominant factors controlling non-visual attributes Retain top 5 SVM detection scores Image patch level features vs whole image features Combining predictors from two or more cities Experiment with smaller cities Use fc7 features or finetune CNN
33
Thank You!
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.