2020/2/6
Abstract
A Trainable CNN for Joint Description and Detection of Local Features 先描述再检测
(1)在 large-scale SfM reconstructions数据集上训练
(2)Aachen Day-Night localization dataset
(3)InLoc indoor localization benchmark
(4)image matching 和 3D reconstruction
之前的“先检测后描述”的方法,检测器对大的外观变化不稳定,可通过密集点描述解决,但是效率低下
本文的缺陷:密集描述子仍不够高效,特征点检测不准确;对外观变化稳定
3.method
Feature Description .pngimage.png
image.png
对于特征点检测器,有2个要求:1.该通道是所有通道中的最大者;2.该点是小邻域9个点中的最大者
Hard feature detection.png
<soft local-max>
<channel-wise non-maximum suppression>
image.png
image-level normalization:
在test阶段,为获得尺度不变性,增加图像金字塔,在不同分辨率获取描述子add,add时采用双线性插值,add公式入下:
image.png
4. Jointly optimizing detection and description loss
损失函数将detection阶段考虑到描述子学习过程中
image.png
image.png
image.png
image.png
image.png
4.2 training
MegaDepth dataset [26]
VGG16 architecture conv4 3 layer,
initial learning rate = 10−3,divided by 2 every 10 epochs
256 × 256 crop centered around one correspondence
barch_size=1
100 images
test
增加分辨率,换成空洞卷积......
5.Experimental Evaluation
1.standard image matching task 具有光照和视角变化
HPatches dataset [5]
根据单应性矩阵H计算的重投影误差如果小于一个阈值则认为是匹配的
We vary the threshold and record the mean matching accuracy (MMA) [32] over all pairs, i.e., the average percentage of correct matches per image pair.每一对图正确匹配的百分比
2.3D reconstruction 具有白天和黑夜的视觉变化
对定位要求很高
dataset:Madrid Metropolis, Gendarmenmarkt and Tower of London [63])
local feature evaluation benchmark [51]
指标:number of images and 3D points;
the mean track lengths of the 3D points;
the mean reprojection error;
dense points
3.visual localization具有较弱的纹理
1).Day-Night Visual Localization
Aachen Day-Night dataset [45,47]
2).Indoor Visual Localization.
InLoc dataset [58]
网友评论