DCFNet: Discriminant Correlation Filters Network for Visual Tracking 阅读
该研究者将DCF类比于Siamese架构中的辅助滤波器模块,并基于将网络输出建模为目标位置概率分布的方法来实现反向传播的计算。通过整合特征提取模块与辅助滤波器模块,在统一框架下完成模型训练。
Abstract.
With DCF-based methods, it has increasingly demonstrated that such approaches represent a significant advancement in the field of online object tracking.
基于DCF的跟踪方法已经成为一种主流的跟踪方法。
The feature sets employed by these approaches are either derived from handcrafted features such as HoGs or convolutional feature maps generated independently across different tasks.
然而此类方法主要依赖于人工提取特征(如HOG),并在其他领域中(例如图像分类)基于预训练的CNN模型提取特征。
In this research, we introduce a novel end-to-end network architecture, referred to as DCFNet, which is designed to enable the extraction and analysis of convolutional features alongside the execution of correlation tracking tasks in a unified framework.
本文作者开发了一种高效轻量型网络以提取并学习卷积特征协同完成相关跟踪过程
Specifically, we integrate DCF as a specialized correlation filter layer within a Siamese network and formulate its backpropagation by setting the network output to represent the probability heatmap of object locations.
具体而言,在其提出的框架中,作者将其视为Siamese网络中新增的相关滤波器模块,并通过将网络输出表示为目标位置的概率热图来进行反向传播计算。
As a result, the process of derivation continues within the Fourier frequency domain, and the efficiency characteristic of DCF remains consistent with its theoretical foundation.
由于仍是在傅里叶频域进行推导,因此保留了DCF的效率性质。
This permits our tracker to achieve a frame rate exceeding 60 FPS in testing environments, while gaining a notable improvement in accuracy compared to the KCF method that employs Histograms of Oriented Gradients (HoGs).
在测试期间运行该跟踪器时达到了每秒60帧以上的处理速度,在此过程中其追踪性能各指标均获得明显提升。相较于基于HOG特性的KCF追踪器而言
The extensive testing on OTB-2013, OTB-2015, and VOT2015 benchmark datasets effectively demonstrates the proposed DCFNet tracker's ability to achieve comparable performance to existing top-performing trackers while maintaining a more compact design and significantly higher operational speed.
实验结果验证了算法的精度与速度。
Introduction.
当无法预知目标类别时,在线学习判别信息以达到高精度的需求就显得尤为重要
基于多层次的手工特征,在这一领域的跟踪器(Tracker)和相关滤波器通常是分开设计的。这种情况下表明该类特征能够发挥作用。随后的研究则聚焦于多层次深度特性的融合。尽管这种方法在提升追踪性能方面表现出色,并且不仅在人工设计的基础上进行优化,并且计算负担较重
由于在先前的方法中所使用的特征与相关滤波流程相互独立,在这种情况下,可能导致跟踪效果无法达到最佳状态。
在本文中
Framework

所提出的网络中包含了若干卷积层,在用于离线训练阶段提取先验跟踪信息的同时又构建了一个独立于主网络的专用特征提取模块。随后引入了一个相关滤波器模块,在此过程中将网络输出转换为目标位置的概率热图分布,并实现了在线更新与目标追踪功能。为了优化整体计算效率和减少资源消耗,在设计网络结构时优先选择了轻量化卷积层方案。
参考
DCFNet: Discriminant Correlation Filter Networks in the field of visual tracking (https://arxiv.org/pdf/1704.04057.pdf)
2. https://www.jianshu.com/p/40737e98ae49
3. <>
