School of Electronic Information Engineering , Tianjin University

1. Cross-view Action Recognition via a Transferable Dictionary Pair (TDP) Jia pingping School of Electronic Information Engineering , Tianjin University

2. Outline： Experimental Method 1 Experimental Result 2 School of Electronic Information Engineering , Tianjin University

3. Experimental Method • Method: • a method for view invariant based on sparse representations using a transferable dictionary pair. • Source-view BoVW/ Targe view BoVW • Learn a dictionary pair {Ds,Dt} • represent training videos in the source view and test videos in the target view using the corresponding source and target dictionaries respectively • classifier ： k_NN

4. 2. Learn a dictionary pair {Ds,Dt} Unsupervised: : the feature representations of M videos of shared actions in the source and target views : the sparse representations p source views，one target view

5. 2. Learn a dictionary pair {Ds,Dt} Supervised: : the feature representations of M videos of shared actions in the source and target views : the sparse representations : consist of the ideal "discriminative" sparse codes of shared action videos in both views.

6. 2. Learn a dictionary pair {Ds,Dt} Unsupervised: • Supervised: Formulate the problem of learning a transferable dictionary pair as an optimization problem which can be efficiently solved using the K-SVD algorithm. Ds and Dt are learned by forcing two sets of videos of shared unlabeled/labeled actions in two views to have the same sparse representations.

7. 3. represent training videos in the source view and test videos in the target view using the corresponding source and target dictionaries respectively OMP algorithm S:all BoW of source view; T: orphan /test BoW of target view

8. Experimental Result • * Test dataset ：IXMAS 1、10actors, 11actions, three time per action 2、experiments on all possible pairwise view combinations (20) 3、1500D local (cuboids) and global (shape flow) feature 4、Dictionary size : 50、100、200、300 Sparse : 10、20、30、50 (紫色：unsupervised ; 红色：supervised ) 5、Leave one action class out 6、k-NN classifier k = 1, 2, …, 10

9. 1、Each row corresponds to a source (training) view and each column a target (test) view. 2、The recognition numbers in the bracket are the average recognition accuracies of k-NN without transfer, our unsupervised and supervised approaches respectively.

10. 1、 Each column corresponds to one target view. 2、The first two rows show the recognition accuracies of our unsupervised and supervised approaches.

11. 参考文献 • Jingjing Zheng, Zhoulin Jiang, P.Jonathon Philips and Rama Chellappa. Cross-view Action Recognition via a Transferable Dictionary Pair. School of Electronic Information Engineering , Tianjin University