GIM at TRECVID 2012 The Light Semantic Indexing Task.
Teresa Alonso, Manuel Barrena García, Pablo G. Rodríguez, Antonio J. Polo, Miryam Salas, Marisa Durán, Félix R. Rodríguez, L. Arevalo, Antonio Corral, Mar Ávila, Jorge Martínez-Gil, Francisco Javier Padillo Ruiz, Soledad Martín, Escuela Politécnica · 2012
The Media\t Engineering\t Research\t Group\t (GIM)\t participated\t in\t the\t semantic\t indexing\t task\t at TRECVID 2012. In\t order\t to\t detect\t semantic\t content\t inside\t the\t videos,\t we\t used\t a\t simple\t approach based on\t visual\t features\t including\t color,\t motion\t and\t SURF\t combined\t with\t textual\t metadata\t from annotations. We\t finally\t submitted\t three\t runs\t based\t on\t the\t following\t approaches: L_A_GIM_RUN1_1:\t training\t based\t in\t a\t subset\t of\t shots\t annotated\t by\t the\t collaborative\t annotation. We calculate\t centroids\t of\t every\t visual\t and\t textual\t feature\t for\t every\t distinct\t concept\t and\t compute distances from\t the\t shot\t that\t is\t being\t evaluated\t to\t the\t centroids. Probability\t of\t a\t shot\t representing one concept\t is\t equal\t to\t 100\t minus\t wheighted\t distances:\t color-‐>28%,\t motion-‐>21%,\t SURF-‐>21%, textual features-‐>30%.\t Relationships\t between\t concepts\t add\t or\t subtract\t 10 %\t to\t the\t final probability and\t a\t threshold\t value\t 25 %\t is\t applied\t on\t it\t in\t order\t to\t discard\t shots. L_A_GIM_RUN2_2:\t same\t as\t L_A_GIM_RUN1_1\t but\t using\t another\t subset\t of\t shots\t for\t training\t and\t a threshold value\t 33%. L_A_GIM_RUN3_3:\t similar\t to\t previous\t but\t using\t a\t different\t training\t set\t and\t a\t different\t weighting: color-‐>52.5%,\t motion-‐>17.5%,\t SURF-‐>0%,\t textual\t features-‐>30%,\t and\t a\t threshold\t value\t 0%.