基于多视角低秩表征的短视频多标签学习模型 下载: 887次
Microvideo Multilabel Learning Model Based on Multiview Low-Rank Representation
1 天津大学电气自动化与信息工程学院, 天津 300072
2 北京智芯微电子科技有限公司, 北京 102200
图 & 表
图 1. 本文模型示意图
Fig. 1. Illustration of proposed model
下载图片 查看原文
图 2. 数据集中具有不同标签的视频示例
Fig. 2. Sample video with different labels selected from dataset
下载图片 查看原文
图 3. 收敛性验证图。(a) Zdiff随模型迭代的变化;(b)平均精度随模型迭代的变化
Fig. 3. Convergence verification graphs. (a) Variation of Zdiff with model iteration; (b) variation of average precision with model iteration
下载图片 查看原文
图 4. 不同参数对于平均精度的影响。(a) λ2对于平均精度的影响;(b) λ4对于平均精度的影响
Fig. 4. Effect of different parameters on average precision. (a) Effect of λ2 on average precision; (b) effect of λ4 on average precision
下载图片 查看原文
图 5. 标签相关性矩阵对比。(a)归一化真实标签相关性矩阵;(b)归一化相关性矩阵
Fig. 5. Label correlation matrix comparison. (a) Normalized correlation matrix for true label; (b) normalized correlation matrix after the iteration
下载图片 查看原文
表 1消融实验结果
Table1. Ablation experiment results
Evaluation metrics | No LR | No LF | No LC |
---|
Average precision difference | -0.0508 | -0.2423 | -0.0172 | Hamming loss difference | 0.0013 | 0.0029 | 0.0005 | Ranking loss difference | 0.0079 | 0.0521 | 0.0049 | Coverage difference | 0.3492 | 4.6914 | 0.2851 | One-error difference | 0.0175 | 0.1482 | 0.0230 |
|
查看原文
表 2不同算法的性能对比
Table2. Performance comparison of different algorithms
Method | Average precision | Hamming loss | Ranking loss | Coverage | One-error |
---|
DNMF | 0.4673±0.0063 | 0.0154±0.0001 | 0.1077±0.0082 | 8.3853±0.1621 | 0.6487±0.0082 | LRR | 0.5489±0.0057 | 0.0154±0.0001 | 0.0991±0.0051 | 8.4056±0.1803 | 0.3039±0.0057 | GLOCAL | 0.7527±0.0064 | 0.0133±0.0020 | 0.0515±0.0015 | 3.9943±0.1056 | 0.2457±0.0032 | MLKNN | 0.7843±0.0053 | 0.0134±0.0001 | 0.0476±0.0058 | 4.0204±0.1874 | 0.3087±0.0058 | Googlenet | 0.6676±0.0044 | 0.0176±0.0002 | 0.4349±0.0066 | 4.5680±0.0600 | 0.4349±0.0066 | C3D | 0.7149±0.0089 | 0.0146±0.0003 | 0.3694±0.0028 | 3.9041±0.2033 | 0.3694±0.0088 | C2AE | 0.8013±0.0022 | 0.0128±0.0001 | 0.0481±0.0041 | 3.6942±0.1471 | 0.2381±0.0026 | Proposed | 0.8055±0.0028 | 0.0128±0.0001 | 0.0432±0.0023 | 3.6732±0.1274 | 0.2561±0.0065 |
|
查看原文
吕卫, 李德盛, 谭浪, 井佩光, 苏育挺. 基于多视角低秩表征的短视频多标签学习模型[J]. 激光与光电子学进展, 2020, 57(22): 221012. Wei Lü, Desheng Li, Lang Tan, Peiguang Jing, Yuting Su. Microvideo Multilabel Learning Model Based on Multiview Low-Rank Representation[J]. Laser & Optoelectronics Progress, 2020, 57(22): 221012.