CVPR-2018
商汤科技
文章目录
1 Background and Motivation
单目标跟踪的难点: illumination, deformation, occlusion and motion
基于深度学习的单目标跟踪精度越来越高,但是 real-time speed 还不足
作者在 SiamFC(【SiamFC】《Fully-Convolutional Siamese Networks for Object Tracking》(ECCV 2016 Workshops)) 的基础上引入 Faster RPN(【Faster RCNN】《Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks》(NIPS-2015)),提出 Siamese Region Proposal Network(SiamRPN)
让跟踪框更加的准确,并且省去多尺度测试耗费的时间
2 Related Work
- Trackers based on Siamese network structure
- RPN in detection
- One-shot learning
- Bayesian statistics based
- meta-learning approaches
3 Advantages / Contributions
-
提出 SiamRPN,which is end-to-end trained off-line with large-scale image pairs for the tracking taskf
-
irst work to formulate online tracking task as one-shot detection
-
又快又强,achieves leading performance in VOT2015, VOT2016 and VOT2017 real-time challenges with the speed of 160 FPS
4 Method
Siamese subnetwork for feature extraction and region proposal subnetwork including the classification branch and regression branch for proposal generation.
template 产生的模板通道数也是 256,correlation operator 卷积核大小 4 x 4 x 2k x 256 or 4 x 4 x 4k x 256,
(1)Siamese feature extraction subnetwork
-
template branch,输入 z z z,网络 φ ( z ) \varphi(z) φ(z)
-
detection branch,输出 x x x, 网络 φ ( x ) \varphi(x) φ(x)
(2)Region proposal subnetwork
-
pair-wise correlation section
φ ( z ) \varphi(z) φ(z) -> [ φ ( z ) ] c l s [\varphi(z)]_{cls} [φ(z)]cls