卫星图像中的车辆分析--A Large Contextual Dataset for Classification, Detection and Counting of Cars
A Large Contextual Dataset for Classification, Detection and Counting of Cars with Deep Learning
ECCV2016
https://gdo-datasci.ucllnl.org/cowc/
本文针对卫星图像中的车辆分析建立了一个新的数据库:Cars Overhead with Context (COWC),然后使用几个 CNN网络对该数据库进行了分析:主要是分类、检测、计数
首先来看看这个新的数据库 Cars Overhead with Context (COWC)
数据库含有 32716个不同的车,来自6个不同的图像库,图像覆盖的区域包括:Toronto Canada [5], Selwyn New Zealand [6], Potsdam [7] and Vaihingen Germany [8], Columbus [9] and Utah [4] United States。
我们的数据库还标记了 58247个有用的负样本,这些样本和正样本比较相似,难以区分,Examples of these are boats, trailers, bushes and A/C units
context is included around targets. Context can help tell us something may not be a car (is sitting in a pond?) or confirm it is a car (between other cars, on a road).
我们对输入图像做了一个归一化,不用考虑车辆的尺度问题。 standardized to 15cm per pixel at ground level from their original resolutions. This makes cars range in size from 24 to 48 pixels。 车辆在图像中的尺寸是 24-48像素之间。有灰度图像,也有彩色图像。
quality, appearance or rotation 这些都是不可控的,需要通过算法来解决
图像是像素级标记的,每个车在其中心点标记一个 dot
The image set is annotated by single pixel points. All cars in the annotated images have a dot placed on their center
对 occlusions, Large trucks, Vans and pickups 做了相应的约定。
我们从卫星图像中间隔的裁出图像块分别作为训练图像和测试图像
测试场景
这里我们对新的数据库上完成三个任务:
1)two-class classifier,即判断图像块中有无车辆
2) detection and localization
3) vehicle counting 这里没有密度图,走检测计数的路线
4 Classification and Detection
设计了一个新的网络结构
我们从卫星图像中裁出 256 × 256 大小的图像块
a set of 308,988 training patches and 79,447 testing patches
4.1 Does Context Help?
从上面可以看出,context 增加到一定之后,性能就下降了。
4.2 Detection
5 Counting
我们是对卫星图像分块计数的。
5.2 Counting Efficiency