PicSOM Experiments in TRECVID 2014

Our experiments in TRECVID 2014 include successful participation in the Semantic Indexing (SIN) task and unsuccessful participation in the Multimedia Event Detection (MED) and Multimedia Event Recounting (MER) tasks. In semantic indexing, we participated in the main task only. We extended our last year’s set of features with SIFT descriptors encoded with Fisher vectors and VLAD, and a total of 24 features based on convolutional neural network (CNN) activations. We also utilized hard negative mining to to acquire more relevant negative examples. We submitted the following four runs: • 4 MUMINPAPPAN: Baseline run matching the best PicSOM SIN submission in TRECVID 2013 • 3 HATTIFNATTAR: Run based on CNN features only, also including hard negative mining • 2 SNUSMUMRIKEN: Run with Fisher vector and VLAD features and the set of 24 CNN features included • 1 MÅRRAN: Run combining all features and hard negative mining The run 1 MÅRRAN obtained the highest MXIAP score of 0.2880. In the Multimedia Event Detection and Recounting task we tried to participate in the MED14-EvalFull search task, but failed.