Crowdsourced Automatic Zoom and Scroll for Video Retargeting Axel Carlier, Wei Tsang Ooi NUS (Singapore) Vincent Charvillat, Romulus Grigoras, Geraldine Morin IRIT (Toulouse, France) 1
iPhone 4 Video retargeting: Making a large video fit into a smaller screen and available with network capacities. 2
One simple way: Scale down the video to the resolution of the screen 3
Example 4
Drawback: Important details may not be visible 5
What makes a good video retargeting ? - Good comprehension of the video content - The video is aesthetically satisfying 6
Important region Zoom 7
Idea : Zoom and scroll to show only interesting regions of the video 8 8
Problem: how to find the region automatically ? 9
Approaches using Content Analysis Saliency map Motion detection Too many regions Liu, Gleicher MM 06 Avidan, Shamir Commun. ACM 09 10 10
Speech Recognition + Natural Language processing Object recognition
Our idea: Crowdsourcing Identifying regions of interest by gathering implicit input from users. 12 12
Viewer Retargeted traces video Passive Active Viewers Viewers
Use zoomable video 14 14
Interface 15 15
Example of user interaction 16 16
Crowdsourcing ● Tutorial: how to use the interface ? ● Magic videos : – HD Videos : 1920 × 1080 pixels – Fixed camera – Obvious ROIs : magician's hands, cards, dice... ● Between 7 and 12 viewers for each video ● 11,183 interaction events logged 17 17
Automatic Generation of Retargeted Video Heatmap ROIs Shot 2 Shots Shot 1 Final Video 18 18
Building Heatmaps 19 19
Building Heatmaps 20 20
Building Heatmaps 21 21
Analyzing Heatmaps Here draw gaussians in 3d with matlab GMM (Gaussian Mixture model) -K -w i Mean Shift -m i - ∑ i 22 22
Finding Modes Mean-Shift: Clustering algorithm (Comaniciu, ICCV 02) 23 23
Determining ROI size Minimum Covariance Determinant (MCD) 24 24
Building a ROI Dynamics Graph
Cutting the graph into shots
Shots selection Shot 1 Shot 3 Shot 2
Result video 28 28
Integrating Reframing techniques ● Bottom-up reframing ● Type of shot: fixed, zooming or dolly ● Shot level: stabilization according to its type ● Inter-shot level: transitions and reestablishing shots 29 29
Dolly shot Shot 3 30 30
Shots 31 31
Shot stabilization 32 32
Transitions 33 33
Reestablishing shots 34 34
Final Result 35 35
Results validation ● 3 poor videos: ● User interaction ( user ) ● Retargeted version without reframing techniques ( noRT ) ● Original version scaled down ( nozoom ) ● Retargeted version with reframing techniques ( crowdsourced ) ● Ground truth ( expert ) 36 36
Ground Truth 37 37
Protocol ● 48 participants divided into 3 categories ● User – crowdsourced – expert (18) ● NoRT – crowdsourced – expert (18) ● Nozoom – crowdsourced – expert (12) ● 3 questions were asked to the participants 38 38
Rate the video editing of the video 5 4 3 2 1 0 39 39 NoRT = retargeted version without reframing techniques
Is the video editing reasonable ? 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 40 40
Does the video manage to convey important information ? 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 41 41
Summary -Gather implicit input from users -No content analysis - In our examples : less than 12 viewers are enough to detect ROIs
Future work ● Explore alternative methods for intermediary steps: ● Modelling heatmaps not as a GMM ● Adding cinematographic rules ● Classify users into different profiles and generate a retargeted video for each profile 43 43
Questions ? 44 44
Results 45 45
Using aesthetics Liu, Chen, Wolf and Cohen-Or. Optimizing Photo Composition, Computer Graphic Forum Luo, Yi wen and Tang, Xiaoou. Photo and Video Quality Evaluation : Focusing on the Subject, ECCV 08 46
MCD Covariance 47
MCD Covariance 48
Gym Video Retargeted 49
Crowdsourcing Shamma, Shaw, Shafton, Liu. Watch what I watch, MIR 07 50
Overview ● Video retargeting ● Zoomable video ● Finding users' interests ● Creating shots ● Integrating reframing techniques ● Results validation 54 54
55 55 http://en.wikipedia.org/wiki/File:UHDV.svg
960 x 640 iPhone 4 56 56 http://en.wikipedia.org/wiki/File:UHDV.svg
57 57
58 58
Approaches using Content Analysis Saliency map Motion detection Liu, Gleicher MM 06 Avidan, Shamir Commun. ACM 09 59 59
Crowdsourcing Heatmap Hotspots Shot 2 Shots Shot 1 Final Video 60 60
Creating Heatmaps ● Modelization of ROIs as a GMM (Gaussian Mixture Model) 61 61
Creating Heatmaps ● Modelization of ROIs as a GMM (Gaussian Mixture Model) 62 62
Building a ROI dynamics graph 63 63
Minimal spanning tree 64 64
Cutting the tree into shots 65 65
Shots selection Shots are selected according to their popularity: 66 66
Recommend
More recommend