Waseda_Meisei at TRECVID 2017: Ad-hoc Video Search
Waseda_Meisei at TRECVID 2017: Ad-hoc Video Search
复制标题
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
K. Ueki;Koji Hirakawa;Kotaro Kikuchi;Tetsuji Ogawa;Tetsunori Kobayashi
中科院分区：
文献类型：
--
作者：
K. Ueki;Koji Hirakawa;Kotaro Kikuchi;Tetsuji Ogawa;Tetsunori Kobayashi
． 早稻田美成团队参加了TRECVID 2017 Ad-hoc视频搜索（AVS）任务[1]。对于今年的AVS任务，我们提交了手动辅助和全自动运行。我们的方法采用以下处理步骤：使用预训练的卷积神经网络（cnn）和支持向量机（svm）构建大型语义概念库，计算所有测试视频的每个概念得分（IACC 3），根据给定的查询短语手动或自动提取多个搜索关键字，并将语义概念得分组合以获得（cid:12）最终搜索结果。我们最好的人工辅助运行达到了21.6%的平均精度（mAP），在所有提交的运行中排名最高。我们最好的全自动跑取得了15.9%的mAP，在所有参与者中排名第二。
. The Waseda Meisei team participated in the TRECVID 2017 Ad-hoc Video Search (AVS) task [1]. For this year’s AVS task, we submitted both manually assisted and fully automatic runs. Our approach used the following processing steps: building a large semantic concept bank using pre-trained convolutional neural networks (CNNs) and support vector machines (SVMs), calculating each concept score for all test videos (IACC 3), manually or automatically extracting several search keywords based on the given query phrases, and combining the semantic concept scores to obtain the (cid:12)nal search result. Our best manually assisted run achieved a mean average precision (mAP) of 21.6%, which ranked the highest among all the submitted runs. Our best fully automatic run achieved a mAP of 15.9%, which ranked second among all participants.