基于中心点搜索的无锚框全卷积孪生跟踪器

为解决孪生网络跟踪器鲁棒性差的问题, 重新设计了孪生网络跟踪器的分类与回归分支, 提出一种基于像素上直接预测方式的高鲁棒性跟踪算法—无锚框全卷积孪生跟踪器(Anchor-free fully convolutional siamese tracker, AFST). 目前高性能的跟踪算法, 如SiamRPN、SiamRPN++、CRPN都是基于预定义的锚框进行分类和目标框回归. 与之相反, 提出的AFST则是直接在每个像素上进行分类和预测目标框. 通过去掉锚框, 大大简化了分类任务和回归任务的复杂程度, 并消除了锚框和目标误匹配问题. 在训练中, 还进一步添加了同类不同实例的图像对, 从而引入了相似语义干扰物, 使得网络的训练更加充分. 在VOT2016、GOT-10k、OTB2015三个公开的基准数据集上的实验表明, 与现有的跟踪算法对比, AFST达到了先进的性能.

AFST: Anchor-free Fully Convolutional Siamese Tracker With Searching Center Point

In order to solve the problem of poor robustness of siamese trackers, this paper redesigns the classification and regression branches, and proposes a high robustness siamese tracker AFST (Anchor-free fully convolutional siamese tracker) based on direct prediction on pixels. Current high-performance object tracker, such as SiamRPN, SiamRPN++, CRPN, are based on predefined anchor boxes for classification and regression. On the contrary, the proposed AFST is to directly classify and predict the target box on each pixel. By removing the anchor, this paper greatly simplifies the complexity of classification task and regression task, and eliminates the problem of mismatching between anchor and target. In the training, we have further added image pairs of different instances of the same kind, thereby introducing similar semantic interferers, making the network training more adequate. Experiments on three open benchmarks datasets, VOT2016, GOT-10k and OTB2015, show that AFST achieves advanced performance compared with existing tracking algorithms.