Fudan at TRECVID 2015: Adaptive Feature Fusion for Multimedia Event Detection in Videos

TRECVID 2015 [4] Multimedia Event Detection (MED) is an interesting and challenging task on the detection of high level complex events in Internet videos [1]. In this notebook paper, we present an overview of our system, focusing on combining multiple feature representations to improve the performance. Specifically, with the outputs of the multiple features, we adopt a simple yet e↵ective fusion method to generate the final predictions, where the optimal fusion weights are learned adaptively for each class, and the learning process is regularized by automatically estimated class relationships. Our MED submissions include 5 system runs for the Pre-Specified (PS) sub-task and 2 runs for the AdHoc (AH) sub-task under 010Ex training condition. Very competitive results are obtained, which verify the e↵ectiveness of both deep features and our fusion method. Table 1: A summary of our submissions. Features Fusion AH Run-1 VGG19-fc67, VGG19-20K, FCVID-233, Conventional Fea. Adaptive Run-2 VGG19-20K, FCVID-233, Conventional Fea. Adaptive PS Run-1 VGG19-fc67, VGG19-20K, FCVID-233, Conventional Fea. Adaptive Param-1 Run-2 VGG19-fc67, VGG19-20K, FCVID-233, Conventional Fea. Average Run-3 VGG19-fc67, VGG19-20K, FCVID-233, Conventional Fea. Adaptive Param-2 Run-4 VGG19-20K, FCVID-233, Conventional Fea. Adaptive Run-5 VGG19-fc67, VGG19-20K, Conventional Fea. Adaptive 1. SYSTEM DESCRIPTION Figure 1 gives an overview of our system, consisting of three components, namely feature extraction, classification and adaptive fusion. We briefly introduce each of them in the following.