A system for TRecViD med by MCIS

We designed a simple system for the 2014 TRECVID Multimedia Event Detection [1]. Except the videos provided by NIST, we only used the BVLC Reference CaffeNet model file distributed besides Caffe [2]. Our system follows the standard pipeline and consists two parts: feature extraction and classification. The feature extraction part is implemented by Caffe and the classification is implemented by LIBSVM [3]. Based on the results, we think that the contribution mainly comes from the feature extraction part. We learned that Convolutional Neural Networks (CNN) is a powerfully model and hope that a easy accessible spatio-temporal CNN model for videos will be available soon.