Event detection in consumer videos using GMM supervectors and SVMs

EURASIP Journal on Image and Video Processing

Table 3 Comparison of STIP, DSTID, and CC-DSTIP

				STIP	STIP	DSTIP	STIP
Event	STIP	DSTIP	CC-DSTIP	+	+	+	+DSTIP
				DSTIP	CC-DSTIP	CC-DSTIP	+CC-DSTIP
Birthday party	0.829	0.853	0.873 (+0.020)	0.839	0.807	0.831	0.820
Changing a vehicle tire	0.834	0.844	0.777 (−0.067)	0.810	0.760	0.783	0.762
Flash mob gathering	0.441	0.500	0.455 (−0.045)	0.436	0.388	0.435	0.387
Getting a vehicle unstuck	0.697	0.683	0.664 (−0.019)	0.664	0.619	0.638	0.608
Grooming an animal	0.785	0.842	0.859 (+0.017)	0.778	0.773	0.782	0.754
Making a sandwich	0.823	0.876	0.873 (+0.003)	0.811	0.785	0.836	0.767
Parade	0.610	0.597	0.673 (+0.076)	0.580	0.576	0.594	0.556
Parkour	0.425	0.497	0.449 (−0.048)	0.444	0.413	0.468	0.446
Repairing an appliance	0.568	0.594	0.533 (−0.061)	0.549	0.537	0.565	0.541
Working on a sewing project	0.760	0.777	0.781 (+0.004)	0.722	0.688	0.762	0.722
Mean	0.677	0.706	0.694 (−0.012)	0.663	0.635	0.669	0.636

Minimum NDCs of (1) three motion features, (2) combination of two of them, and (3) combination of three of them for each TRECVID2011 MED events are reported. The italic figures indicate the feature with the best performance in each category. Figures inside the parenthesis indicate the difference between CC-DSTIP and DSTIP.