FOCUS: Clustering Crowdsourced Videos by Line-of-Sight Puneet Jain, Justin Manweiler, Arup Acharya, and Kirk Beaty Clustered by shared subject CHALLENGES CAN IMAGE PROCESSING SOLVE THIS PROBLEM? Camera 1 Camera 2 LOGICAL similarity does not imply VISUAL similarity Camera 3 Camera 4 5 VISUAL similarity does not imply LOGICAL similarity 6 CAN SMARTPHONE SENSING SOLVE THIS PROBLEM? Why not triangulate? Sensors are noisy, hard to distinguish subjects… GPS-COMPASS Line-of-Sight INSIGHT easy to identify hard to identify Don’t need to visually identify actual SUBJECT, can use background as PROXY Simplifying Insight 1 same basic structure persists Don’t need to directly match videos, can compare all to a predefined visual MODEL Simplifying Insight 2 Light-of-sight (triangulation) is almost enough, just not via sensing (alone) Simplifying Insight 3 FOCUS Fast Optical Clustering of live User Streams Vision Sensing Cloud Clustered Videos Video Extraction Video Streams (Android, iOS, etc.) Hadoop/HDFS Failover, elasticity Image processing Computer vision FOCUS Cloud Video Analytics Watching Live home: 2 away: 1 Change Angle Change Focus Users Select & Watch Organized Streams Clustered Videos Video Extraction Hadoop/HDFS Failover, elasticity Image processing Computer vision FOCUS Cloud Video Analytics pre-defined reference “model” Watching Live home: 2 away: 1 Change Angle Change Focus Users Select & Watch Organized Streams keypoint z extraction multi-view reconstruction z estimates camera POSE and content in field-of-view Multi-view Stereo Reconstruction Model construction technique based on Photo Tourism: Exploring image collections in 3D Snavely et al., SIGGRAPH 2006 17 Visualizing Camera Pose keypoint z extraction multi-view reconstruction z frame-by-frame video to model alignment z z inputs sensory • Given a pre-defined 3D, align incoming video frames to the model • Also known as camera pose estimation 19 keypoint z extraction multi-view reconstruction z integration of sensoryz inputs Gyroscope, provides “diff” from vision initial position Filesize ≈ 1/Blur 0 1 2 Gyroscope Sampled Frame 3 4 t-1 t-2 20 keypoint z extraction multi-view reconstruction z pairwise model image analysis z Field-of-view Using POSE + model POINT CLOUD, FOCUS geometrically identifies the set of model points in background of view 21 keypoint z extraction multi-view reconstruction z pairwise model image analysis z 3 Similarity between image 1 & 3 = 13 Similarity between image 1 & 2 = 18 2 1 Finding the similarity across videos as size of point cloud set intersection 22 Clustering “similar” videos Similarity Score 1 Application of Modularity Maximization 2 2 high modularity implies: 3 3 1 • high correlation among the members of a cluster • minor correlation with the members of other clusters RESULTS Collegiate Football Stadium • Stadium 33K seats 56K maximum attendance • Model: 190K points 412 images (2896 x 1944 resolution) • Android App on Samsung Galaxy Nexus, S3 • 325 videos captured 15-30 seconds each 25 Line-of-Sight Accuracy (visual) 26 Line-of-Sight Accuracy GPS/Compass LOS estimation is <260 meters for the same percentage In >80% of the cases, Line-of-sight estimation is off by < 40 meters 27 FOCUS Performance 75% true positives Trigger GPS/Compass failover techniques 28 Natural Questions • What if 3D model is not available? – Online model generation from first few uploads • Stadiums look very different on a game day? – Rigid structures in the background persists • Where it won’t work? – Natural or dynamic environment are hard Conclusion • Computer vision and image processing are often computation hungry, restricting real-time deployment • Mobile Sensing is a powerful metadata, can often reduce computation burden • Computer vision + Mobile Sensing + Geometry, along with right set of BigData tools, can enable many real-time applications • FOCUS, displays one such fusion, a ripe area for further research Thank You http://cs.duke.edu/~puneet