Robust View-Invariant Representation for Classification and Retrieval of Image and Video Data

for their guidance, advice and encouragement throughout my graduate study. Their infinite passion and serious attitude in scientific research will benefit me more in my future life. I also want to thank my doctoral dissertation defense committee, Profs. Shmuel Friedland, Daniela Tuninetti and Milos Zefran for their help and suggestions. Finally, I would like to thank my parents, for their support at all times. 2 (a) Visual illustration for trajectories normalization for of the " come " class from ASL dataset. The " + " curves are original trajectories and • curves are normalized trajectories. (b) Comparison of retrieval time with three approaches from ASL and CAVIAR dataset, X axis stands for the segmentation length, Y axis stands for the retrieval time (sec).. 30 3 Visual illustration for retrieval results with 20 classes with motion trajec-tories from CAVIAR dataset for the motion events " chase " and " shopping and leave " for fixed cameras from unknown views.. .. .. .. .. . 31 4 (a) Accuracy values for the classification problem on increasing number of classes, (b) Accuracy values for the classification problem on increasing number of trajectories in a class. (c) The comparison with Precision-Recall Curve for retrieval with 20 classes for perfect trajectories.. .. . 31 5 (a) PR Curve for classification with 40 classes with true and noisy tra-jectories with and without segmentation when the noise variance is 0.5. (b) PR Curve for indexing and retrieval with 20 classes with perfect and noisy trajectories with and without segmentation when the noise 6 The first row are the visual illustrations for the query and the top three retrieval of the video frames with overlapping segmented approaches in CAVIAR dataset, the second row are the corresponding first 16 non-zero null space invariants with global motion trajectories, the third row are the corresponding first 16 non-zero null space invariants with overlapping segmented motion trajectories.. 7 The first row are the visual illustrations for the query and the top three retrieval of the video frames with overlapping segmented approaches in CAVIAR dataset, the second row are the corresponding first 16 non-zero null space invariants with overlapping segmented trajectories , the third row are the corresponding first null space invariants with non-overlapping segmented motion trajectories.. FIGURE PAGE 8 Three sample trajectories from ASL dataset. For each trajectory, another two versions of noisy trajectories with arbitrary moving cameras are also shown. Original …