PKU_ICST at TRECVID 2018: Instance Search Task

We participated in all two types of instance search (INS) task in TRECVID 2015: automatic search and interactive search. This paper presents our approaches and results. In this task, we mainly focused on exploring the effective feature representation, feature matching and re-ranking algorithm. In this year, we also tried to use Deep Neural Networks (DNN) to improve the results. In feature representation, we extracted two kinds of features: (1) Bag-of-Words (BoW) feature based on Approximate K-means (AKM) and (2) DNN feature based on Convolutional Neural Networks (CNN). In feature matching, we adopted different ranking methods to different features: (1) For the AKM-based BoW feature, we used cosine distance to calculate the similarity between each query topic and each shot; (2) For the DNN feature, multi-bag SVM (MBSVM) was adopted since it can make full use of all query examples. Moreover, we conducted keypoint matching algorithm on the top ranked results. It was effective yet efficient since only top ranked results were considered. In re-ranking stage, we further incorporated transcripts into our framework to explore the context information. The official evaluations showed that our team is ranked 1 st on both automatic search and interactive search.