其他分享
首页 > 其他分享> > 目标跟踪数据集整理(五)---NfS

目标跟踪数据集整理(五)---NfS

作者:互联网

文章目录


Need for Speed: A Benchmark for Higher Frame Rate Object Tracking 2017
从文章的题目可以看出关键字“higher frame rate(240fps)”,这个数据集包含了100个数据集,每个视频序列都是手工标注,9个 挑战属性。75 videos using the iPhone 6 and the iPad Pro,25 sequences from YouTube.

motivation: 现有的大多数数据集的帧速率是30fps,但是现在的一些消费型设备如相机、手机、平板和机器人等都需要较高的帧速率 (240 FPS now being standard on many smart phones, tablets, drones, etc.).Specifically, we want to explore
that if we actually increase the frame rate - thus reducing the amount of appearance variation per frame - could we get
away with substantially simpler tracking algorithms (from a computational perspective)?

我们工作一个令人兴奋的结果是 如果可以获得足够高的帧速率,使用廉价的手工特征(HOG)的CF与最先进的深度特征相比,可以实现非常有竞争力甚至更高的精度和较优的计算效率(也就是说,在较高的帧速率下,简单的跟踪器如CF超过复杂的基于深度的方法)为甚呢?
for higher frame rate video there is less appearance change among consecutive frames,which can be efficiently modeled by updating the tracking model at each frame even using simple hand-craffed features

Related Work

Tracking Datasets

OTB50 /OTB100: OTB50 is a subset of OTB100,11 attributes
TC128: 128 videos,specifically designed for the evaluation of color-enhanced trackers,11 attributes
VOT14 and VOT15: 25 and 30 videos,all videos are labelled with rotated bounding boxes-rather than upright ones-with per frame attribute annotation.
ALOV300: 314 videos,14 attributes
UAV123: 128 videos ,115 videos captured by UAV cameras and 8 sequences rendered by a UAV simulator,12 attributes
在这里插入图片描述

Tracking Methods

(1) CF trackers: MOSSE,KCF…showed that learning such trackers in the frequency domain is highly affected by boundary effects,leading to suboptimal performance
learning CF trackers from deep Convolutional Neural Networks features: the major drawback of such CF trackers is their intractable complexity(~0.2fps on CPUs) mainly needed for extracting deep features and computing Fourier transforms on hundreds of feature channels,
(2) deep trackers
two high-speed deep trackers GOTURN and SFC are able to run at 100 FPS and 75 FPS on GPU.drawback: not update/ on the modern CPUs,their speed becomes no more than 3FPS,which is too slow for practical use on devices with limited computational resources.

Evaluation

at high frame rate,the motion blur caused by fast moving object/camera is significantly diminished .This leads to excluding the effect of motion blur in lower frame rate tracking. so it needs to simulate motion blur.在这里插入图片描述
为了保证CF tracke这类方法在高帧速率下的鲁棒性,我们近似的调整学习率,Lnew=1/8 L old。

The result shows that adjusting the learning rates notably improves the accuracy of all the CF trackers.All CF trackers achieve a significant increase in performance when tracking on 240FPS videos.This is because in higher frame rate video,the appearance change between two adjacent frames is very small,which can be effectively learned by per frame CF online adaptaion.
对于深度方法,提高最低的是SFC和MDNet,这些方法离线训练并且不更新或者保存过去外观的一个记忆,因此高帧率的跟踪视频对这些跟踪器提供较低的提升。

在CUP 上所有的CF 跟踪器相比于所有的深度跟踪器获得了更高的速度,因为它们较浅的架构和在傅里叶域有效的计算。
在GPU 上深度跟踪器比在CPU上快了好多。
最后放一张图,在不同的帧率下不同的挑战属性下,不同跟踪器的表现:在这里插入图片描述

标签:videos,trackers,frame,CF,跟踪器,rate,NfS,跟踪,整理
来源: https://blog.csdn.net/xwmwanjy666/article/details/99288396