WHU-NERCMS at TRECVID2019: Instance Search Task

Longxiang Jiang, Jingyao Yang, Erxuan Guo, Fan Xia, Ruxing Meng, Jingfeng Luo, Xiangyu Li, Xinyi Yan, Zengmin Xu, Chao Liang
2019 TREC Video Retrieval Evaluation  
This paper presents the framework and results of our participation in the TRECVID 2019 Instance Search task. This year the instance search task aims at retrieving specific persons doing specific actions. In fact, this retrieval target can be divided into two parts, that is, person and action retrieval. So we tried two schemes: step retrieval and overall retrieval. On the one hand, we try to get the final retrieval results by retrieving the peolple and actions respectively and fusing them later.
more » ... The advantage of this method lies in the independence of retrieval and the simplicity of later fusing. On the other hand, we try to get the final retrieval results by retrieving, tracking specific people and retrieving their actions. The advantage of this method is that it adopts targeted action retrieval for specific people to ensure the integrity of retrieval. The two retrieval schemes are complementary. With the proposed methods, our first scheme get 0.016 mAP and our second scheme get 0.017 mAP in automatic task according the evaluation, ranks 4 among all instance search teams.
dblp:conf/trecvid/JiangYGXMLLYXL19 fatcat:a7xyov4cn5gf7gnyf5l6tduk3a