WasedA at TRECVID 2015: Semantic indexing

Waseda participated in the TRECVID 2015 Semantic Indexing (SIN) task [6]. For the SIN task, our approach used the following processing pipelines: feature extraction using several deep convolutional neural networks (CNNs); classification of the presence or absence of a detection target by support vector machines (SVMs); and fusion of multiple score outputs. In order to improve the performance of semantic video indexing, we employed the following techniques: utilizing multiple evidences observed in each video and compressing them into a fixed-length vector; introducing gradient and motion features to CNNs; enriching variations of the training and the testing sets; and extracting features from several CNNs trained with various large-scale datasets. Through these techniques, our best run achieved a mean Average Precision (mAP) of 30.9%. This was ranked 2nd among all the participants.