Command Palette
Search for a command to run...
Qiang Wang; Li Zhang; Luca Bertinetto; Weiming Hu; Philip H.S. Torr
摘要
本文展示了如何通过一种简单的方法实现实时的视觉目标跟踪和半监督视频对象分割。我们提出的方法称为SiamMask，通过在流行的全卷积Siamese目标跟踪方法的离线训练过程中增加一个二值分割任务来改进其损失函数。训练完成后，SiamMask仅依赖于单个边界框初始化，并在线运行，以每秒55帧的速度生成类别无关的对象分割掩码和旋转边界框。尽管该方法简单、灵活且速度快，但我们的策略在VOT-2018实时跟踪器中建立了新的最先进水平，同时在DAVIS-2016和DAVIS-2017数据集上展示了具有竞争力的性能和最佳速度，用于半监督视频对象分割任务。项目网站为：http://www.robots.ox.ac.uk/~qwang/SiamMask。
代码仓库
shallowtoil/DROL
pytorch
GitHub 中提及
ezelikman/anonymal
pytorch
GitHub 中提及
foolwood/SiamMask
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| semi-supervised-video-object-segmentation-on-1 | SiamMask | F-measure (Decay): 22.4 F-measure (Mean): 45.8 F-measure (Recall): 45.3 Ju0026F: 43.2 Jaccard (Decay): 21.9 Jaccard (Mean): 40.6 Jaccard (Recall): 44.5 |
| video-object-tracking-on-nv-vot211 | SiamMask | AUC: 35.14 Precision: 46.49 |
| visual-object-tracking-on-davis-2016 | SiamMask | F-measure (Decay): 2.1 F-measure (Mean): 67.8 F-measure (Recall): 79.8 Ju0026F: 69.75 Jaccard (Decay): 3.0 Jaccard (Mean): 71.7 Jaccard (Recall): 86.8 |
| visual-object-tracking-on-davis-2017 | SiamMask | F-measure (Decay): 20.9 F-measure (Mean): 58.5 F-measure (Recall): 67.5 Ju0026F: 56.4 Jaccard (Decay): 19.3 Jaccard (Mean): 54.3 Jaccard (Recall): 62.8 |
| visual-object-tracking-on-vot201718 | SiamMask | Expected Average Overlap (EAO): 0.380 |
| visual-object-tracking-on-youtube-vos | SiamMask | F-Measure (Seen): 58.2 F-Measure (Unseen): 47.7 Jaccard (Seen): 54.3 Jaccard (Unseen): 45.1 O (Average of Measures): 52.8 |