Clustering crowdsourced videos by line-of-sight

摘要

A method for clustering images includes acquiring initial image data including a scene of interest. A 3D model is constructed of the scene of interest based on the acquired initial image data. Additional image data including the scene of interest is acquired. The additional image data is fitted to the 3D model. A line-of-sight of the additional image data is estimated based on the fitting to the 3D model. The additional image data is clustered according to the estimated line-of-sight.

主权项

1. A method for clustering images, comprising:
acquiring initial image data including a scene of interest; constructing a 3D model of the scene of interest based on the acquired initial image data; acquiring a plurality of additional images including the scene of interest; fitting each of the plurality of additional images to the 3D model; estimating a line-of-sight for each of the plurality of additional images based on the corresponding fitting to the 3D model; and clustering the plurality of additional images into a plurality of clusters according to the estimated line-of-sight of each of the plurality of additional images, wherein the clustering of the plurality of additional images into the plurality of clusters according to the estimated line-of-sight includes: estimating a 3D ray for each estimated line-of-sight; calculating a shortest distance between each pair of estimated 3D rays; and clustering the 3D rays according to the shortest distances, wherein calculating a shortest distance between each pair of estimated 3D rays includes analyzing only pairs of 3D rays that converge and calculating a distance of a line segment that connects an origin of a first 3D ray of each pair of 3D rays and intersects a second 3D ray of each pair of 3D rays perpendicularly.