Clustering Photos to Improve Visualization of Collections
One of the results of requirement analysis (done in aceMedia) was that clustering visually similar photos without disturbing the timeline would be very helpful while searching and navigating. This thesis determines an appropriate clustering algorithm to cluster visually similar photos without disturbing the timeline.
Considering the results of requirement analysis, three clustering algorithms were considered; time-based, time & visual similarity-based, and visual similarity-based. The mean of Color Histogram of photos was used to determine the visual similarity.
The clusters found by the algorithms were compared with that of the clusters found by users in the same photo collection. F1 metric was used for the comparison. Two of the three considered clustering algorithms (time-based and time & visual similarity-based) produced similar and expected results. To determine the appropriateness and completeness of the clusters found by these two algorithms, users were asked to evaluate the clusters. The results obtained were favorable to both the algorithms. Considering the response time of the algorithms time-based algorithm was found to be an appropriate algorithm.
Click here to download my thesis documentation.
Prof. Dr. Reinhard Oppermann