This manuscript conducts a comparison on modern object detection systems in their ability to detect multiple maritime vessel classes. Three highly scoring algorithms from the Pascal VOC Challenge, Histogram of Oriented Gradients by Dalal and Triggs, Exemplar-SVM by Malisiewicz, and Latent-SVM with Deformable Part Models by Felzenszwalb, were compared to determine performance of recognition within a specific category rather than the general classes from the original challenge. In all cases, the histogram of oriented edges was used as the feature set and support vector machines were used for classification. A summary and comparison of the learning algorithms is presented and a new image corpus of maritime vessels was collected. Precision-recall results show improved recognition performance is achieved when accounting for vessel pose. In particular, the deformable part model has the best performance when considering the various components of a maritime vessel.
Comparison of Object Detection Algorithms on Maritime Vessels
2014
12 pages
Report
No indication
English
Algebra, Analysis, Geometry, & Mathematical Logic , Computers, Control & Information Theory , Marine Engineering , Algorithms , Computer vision , Detection , Comparison , Histograms , Naval vessels , Object detection algorithms , Maritime vessels , Svm(support vector machines) , Learning algorithms , Hog(histogram of oriented gradients) , Voc(visual object classes)
Engineering Index Backfile | 1953