【正文】
Hybrid:% Fig. Comparison of classification performance for using motion, static and hybrid features. 2022/8/12 26 Experiments—— YouTube dataset(3/3) Fig. The confusion table for classification using hybrid features. 2022/8/12 27 Conclusions Interest This paper present a systematic framework for recognizing realistic actions from videos “in the wild”. Static features are plementary to motion features. Using Motion cues to prune motion and static features is helpful. Informationtheoretic based divisive clustering reconstruct pact yet discriminative semantic visual vocabularies. 2022/8/12 28 Thank you! 2022/8/12 29 。 Output: visual word clusters Initiate randomly assign the cluster members This is similar to kmeans Two major steps For each cluster ,pute the prior and “centers”. Update clusters : for each ,find the new cluster: X?X?)?|( xCp)|( XCp??()tiitxxx???? ???( | ) ( | )?()tititxx ip C x p C xx???? ??ixtx* ?( ) a r g m in K L ( ( | ) , ( | ) )t j t ji x p C x p C x?2022/8/12 22 Experiments—— KTH dataset static features % motion features % Hybrid features % Static feature: shape information 2022/8/12 23 YouTube dataset b_shooting g_walking t_jumping s_juggling cycling t_swing t_swinging v_spiking diving swinging r_riding 11 categories About 1600 videos 2022/8/12 24 Experiments—— YouTube dataset(1/3) Figure A:Performance parison between system with motion feature pruning and without feature pruning Figure B:Performance parison between system with static feature pruning and without feature pruning Average Accuracy Before pruning : 57% After pruning : % Average Accuracy Before pruning : % After pruning : % 2022/8/12 25 Experiments—— YouTube dataset(2/3) Average accuracy: Motion: %。 Localization Contributions 2022/8/12 12 Motivation(1/3)—— Static Features Why Static Features? In Realistic video, motion features are unreliable due to unpredictable and often unintended camera motion (camera shake). Correlated o