E XEMPLAR -SVM FOR A CTION R ECOGNITION Week 11 Presented by Christina Peterson

C HANGES MADE TO C OMBINED E XEMPLAR -SVM S Multi-Class SVM trained on calibrated exemplar scores rather than raw exemplar-svm scores Ran STIP for Kicking action class to obtain descriptors for more frames

R ECOGNITION A CCURACIES ON UCF S PORTS DATA SET Combined Exemplar-SVMs increased from 67.3% accuracy to 75% accuracy Method Accuracy (%)DivingGolfingKickingLiftingRidingRunningSkating Swing- bench High- swingWalking Rodriguez et al. [1] Yeffet and Wolf [2] Le et al. [4] Wu et al. [6] Action Bank [7] Standard Multiclass-SVMs Combined Exemplar-SVMs

C ONFUSION M ATRIX : C OMBINED E XEMPLAR -SVM DiGoSsSbSkSk RuHoLiKiWaWa Diving Golf Kick Lift Horse-Ride Run Skateboar d Swing- bench Swing-side Walk

C ONCLUSIONS The changes have made performance comparable to Standard Multi-Class SVM The selected exemplar set has a large impact on the accuracy on the test set Improving accuracy would involve manually selecting the best exemplars to represent the action class

R EFERENCES [1] M. D. Rodriguez, J. Ahmed, and M. Shah. Action mach: A spatio-temporal maximum average correlation height filter for action recognition. In CVPR, [2] Yeffet and L. Wolf. Local trinary patterns for human action recognition. In ICCV, [3] H. Wang, M. Ullah, A. Klaser, I. Laptev, and C. Schmid. Evaluation of local spatio-temporal features for action recognition. In BMVC, [4] Q. Le, W. Zou, S. Yeung, and A. Ng. Learning hierarchical invariant spatiotemporal features for action recognition with independent subspace analysis. In CVPR, [5] A. Kovashka and K. Grauman. Learning a hierarchy of discriminative spacetime neighborhood features for human action recognition. InCVPR, [6] X. Wu, D. Xu, L. Duan, and J. Luo. Action recognition using context and appearance distribution features. InCVPR, [7] S. Sadanand and J. J. Corso. Action bank: A high-level representation of activity in video. CVPR, 2012.