Tracking"沦陷”啦，又一个被deep learning拿下来的领域

贾伟中科院合肥

http://arxiv.org/pdf/1510.07945v1.pdf


Learning Multi-Domain Convolutional Neural Networks for Visual Tracking.pdf (3.11 MB, 下载次数: 3980)

2015-11-1 23:24 上传

点击文件名下载附件


We propose a novel visual tracking algorithm based on the representations from adiscriminatively trained Convolutional Neural Network (CNN). Our algorithmpretrains a CNN using a large set of videos with tracking ground truths toobtain a generic target representation. Our network is composed of sharedlayers and multiple branches of domain-specific layers, where domainscorrespond to individual training sequences and each branch is responsible forbinary classification to identify target in each domain. We train each domainin the network iteratively to obtain generic target representations in theshared layers. When tracking a target in a new sequence, we construct a newnetwork by combining the shared layers in the pretrained CNN with a new binaryclassification layer, which is updated online. Online tracking is performed byevaluating the candidate windows randomly sampled around the previous target state.The proposed algorithm illustrates outstanding performance in existing trackingbenchmarks。

邵枭虎中科院

各种被攻陷。

吴悠西北工大

  现在deep learning无所不能了，如果速度能提上去就好了，或者硬件那边再来次大革新