融合扩张卷积网络与SLAM的无监督单目深度估计

戴仁月; 方志军; 高永彬

doi:doi:10.3788/LOP57.061007

激光与光电子学进展, 2020, 57 (6): 061007, 网络出版: 2020-03-06

融合扩张卷积网络与SLAM的无监督单目深度估计下载： 1189次

Unsupervised Monocular Depth Estimation by Fusing Dilated Convolutional Network and SLAM

戴仁月方志军 ^*高永彬

作者单位

上海工程技术大学电子电气工程学院, 上海 201600

图 & 表

图 1. 网络框架示意图

Fig. 1. Illustration of the network framework

下载图片查看原文

图 2. 标准卷积与扩张卷积滤波器对比图。(a)标准卷积滤波器;(b)扩张率为2的扩张卷积滤波器;(c)扩张率为3的扩张卷积滤波器

Fig. 2. Comparison of standard convolution and dilated convolution filters. (a) Standard convolution filter; (b) dilated convolution filter with dilation ratio of 2; (c) dilated convolution filter with dilation ratio of 3

下载图片查看原文

图 3. 扩张卷积与标准卷积的可视化过程对比图。(a)标准卷积可视化过程;(b)扩张率为2的扩张卷积可视化过程;(c)扩张率为3的扩张卷积可视化过程

Fig. 3. Visualization process comparison of dilated convolution and standard convolution. (a) Visualization process of standard convolution; (b) visualization process of dilated convolution with dilation ratio of 2; (c) visualization process of dilated convolution with dilation ratio of 3

下载图片查看原文

图 4. ORB-SLAM算法优化全局相机姿态总体流程

Fig. 4. Flow chart of optimizing global camera pose by ORB-SLAM algorithm

下载图片查看原文

图 5. 三维空间点在图像平面上的投影过程

Fig. 5. Projection process of three-dimensional space points onto the image plane

下载图片查看原文

图 6. 不同损失变化曲线。(a)重建损失;(b)平滑损失;(c)总体损失

Fig. 6. Curves for different losses. (a) Reconstruction loss; (b) smooth loss; (c) total loss

下载图片查看原文

图 7. KITTI Odometry数据集中不同序列的相对相机姿态轨迹。(a) 00;(b) 01;(c) 09;(d) 02;(e) 03;(f) 10

Fig. 7. Camera pose trajectories for different sequences in the KITTI Odometry dataset. (a) 00; (b) 01; (c) 09; (d) 02; (e) 03; (f) 10

下载图片查看原文

图 8. 深度预测的定性比较。(a) RGB输入图像;(b) Garg等^[11]的方法;(c) sfmlearner方法^[4];(d)本文方法;(e) ground truth

Fig. 8. Qualitative comparison of depth prediction. (a) RGB input image; (b) method of Garg et al.^[11]; (c) sfmlearner method^[4]; (d) our method; (e) ground truth

下载图片查看原文

图 9. 深度细节可视化比较。(a)(c)输入图像;(b)(d)输出图像

Fig. 9. Visualization comparison of depth details. (a)(c) Input images; (b)(d) output images

下载图片查看原文

表 1KITTI Odometry数据集09和10序列的RMSE比较

Table1. RMSE comparison of 09 and 10 sequences in the KITTI Odometry dataset

Method	Sequence 09		Sequence 10
Method	t_error /%	r_error per100 m /(°)	t_error /%	r_error per100 m /(°)
Luo et al.^[20]	3.72	1.60	6.06	2.22
Zhou et al.^[4]	18.77	3.21	14.33	3.30
Li et al.^[21]	7.01	3.61	10.63	4.65
Zhanet al.^[13] (Tem)	11.93	3.91	12.45	3.46
Zhan et al.^[13](New YorkUniversitydatasets)	11.92	3.60	12.62	3.43
Ours	1.70	0.50	1.43	0.52

查看原文

表 2深度估计模型的TUM评估结果比较

Table2. Comparison of TUM evaluation results for depth estimation model

Method	Supervised	Data	Error				Accuracy
Method	Supervised	Data	A	S	R	lg R	δ₁ /%	δ₂ /%	δ₃ /%
Method in Ref. [5]	√	KITTI	0.214	1.605	6.563	0.292	67.3	88.4	95.7
Method in Ref. [6]	√	KITTI	0.203	1.548	6.307	0.282	70.2	89.0	95.8
Method in Ref. [7]	√	KITTI	0.202	1.614	6.523	0.275	67.8	89.5	96.5
Method in Ref. [22] (photo)	×	KITTI	0.211	1.980	6.154	0.264	73.2	89.8	95.9
Method in Ref. [22] (photo+ad)	×	KITTI	0.220	1.976	6.340	0.273	70.8	86.7	93.4
Method in Ref. [4]	×	KITTI	0.208	1.768	6.856	0.283	67.8	88.5	95.7
Method in Ref. [4](without explainability masks)	×	KITTI	0.221	2.226	7.527	0.294	67.6	88.5	95.4
Ours	×	KITTI	0.189	1.592	6.432	0.268	71.4	91.1	96.3

查看原文

戴仁月, 方志军, 高永彬. 融合扩张卷积网络与SLAM的无监督单目深度估计[J]. 激光与光电子学进展, 2020, 57(6): 061007. Renyue Dai, Zhijun Fang, Yongbin Gao. Unsupervised Monocular Depth Estimation by Fusing Dilated Convolutional Network and SLAM[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061007.

融合扩张卷积网络与SLAM的无监督单目深度估计下载： 1189次

图 1. 网络框架示意图

Fig. 1. Illustration of the network framework

图 2. 标准卷积与扩张卷积滤波器对比图。(a)标准卷积滤波器;(b)扩张率为2的扩张卷积滤波器;(c)扩张率为3的扩张卷积滤波器

Fig. 2. Comparison of standard convolution and dilated convolution filters. (a) Standard convolution filter; (b) dilated convolution filter with dilation ratio of 2; (c) dilated convolution filter with dilation ratio of 3

图 3. 扩张卷积与标准卷积的可视化过程对比图。(a)标准卷积可视化过程;(b)扩张率为2的扩张卷积可视化过程;(c)扩张率为3的扩张卷积可视化过程

Fig. 3. Visualization process comparison of dilated convolution and standard convolution. (a) Visualization process of standard convolution; (b) visualization process of dilated convolution with dilation ratio of 2; (c) visualization process of dilated convolution with dilation ratio of 3

图 4. ORB-SLAM算法优化全局相机姿态总体流程

Fig. 4. Flow chart of optimizing global camera pose by ORB-SLAM algorithm

图 5. 三维空间点在图像平面上的投影过程

Fig. 5. Projection process of three-dimensional space points onto the image plane

图 6. 不同损失变化曲线。(a)重建损失;(b)平滑损失;(c)总体损失

Fig. 6. Curves for different losses. (a) Reconstruction loss; (b) smooth loss; (c) total loss

图 7. KITTI Odometry数据集中不同序列的相对相机姿态轨迹。(a) 00;(b) 01;(c) 09;(d) 02;(e) 03;(f) 10

Fig. 7. Camera pose trajectories for different sequences in the KITTI Odometry dataset. (a) 00; (b) 01; (c) 09; (d) 02; (e) 03; (f) 10

图 8. 深度预测的定性比较。(a) RGB输入图像;(b) Garg等^[11]的方法;(c) sfmlearner方法^[4];(d)本文方法;(e) ground truth

Fig. 8. Qualitative comparison of depth prediction. (a) RGB input image; (b) method of Garg et al.^[11]; (c) sfmlearner method^[4]; (d) our method; (e) ground truth

图 9. 深度细节可视化比较。(a)(c)输入图像;(b)(d)输出图像

Fig. 9. Visualization comparison of depth details. (a)(c) Input images; (b)(d) output images

表 1KITTI Odometry数据集09和10序列的RMSE比较

Table1. RMSE comparison of 09 and 10 sequences in the KITTI Odometry dataset

表 2深度估计模型的TUM评估结果比较

Table2. Comparison of TUM evaluation results for depth estimation model

关于本站 Cookie 的使用提示

全站搜索

融合扩张卷积网络与SLAM的无监督单目深度估计 下载： 1189次

图 1. 网络框架示意图

Fig. 1. Illustration of the network framework

图 2. 标准卷积与扩张卷积滤波器对比图。(a)标准卷积滤波器;(b)扩张率为2的扩张卷积滤波器;(c)扩张率为3的扩张卷积滤波器

Fig. 2. Comparison of standard convolution and dilated convolution filters. (a) Standard convolution filter; (b) dilated convolution filter with dilation ratio of 2; (c) dilated convolution filter with dilation ratio of 3

图 3. 扩张卷积与标准卷积的可视化过程对比图。(a)标准卷积可视化过程;(b)扩张率为2的扩张卷积可视化过程;(c)扩张率为3的扩张卷积可视化过程

Fig. 3. Visualization process comparison of dilated convolution and standard convolution. (a) Visualization process of standard convolution; (b) visualization process of dilated convolution with dilation ratio of 2; (c) visualization process of dilated convolution with dilation ratio of 3

图 4. ORB-SLAM算法优化全局相机姿态总体流程

Fig. 4. Flow chart of optimizing global camera pose by ORB-SLAM algorithm

图 5. 三维空间点在图像平面上的投影过程

Fig. 5. Projection process of three-dimensional space points onto the image plane

图 6. 不同损失变化曲线。(a)重建损失;(b)平滑损失;(c)总体损失

Fig. 6. Curves for different losses. (a) Reconstruction loss; (b) smooth loss; (c) total loss

图 7. KITTI Odometry数据集中不同序列的相对相机姿态轨迹。(a) 00;(b) 01;(c) 09;(d) 02;(e) 03;(f) 10

Fig. 7. Camera pose trajectories for different sequences in the KITTI Odometry dataset. (a) 00; (b) 01; (c) 09; (d) 02; (e) 03; (f) 10

图 8. 深度预测的定性比较。(a) RGB输入图像;(b) Garg等[11]的方法;(c) sfmlearner方法[4];(d)本文方法;(e) ground truth

Fig. 8. Qualitative comparison of depth prediction. (a) RGB input image; (b) method of Garg et al.[11]; (c) sfmlearner method[4]; (d) our method; (e) ground truth

图 9. 深度细节可视化比较。(a)(c)输入图像;(b)(d)输出图像

Fig. 9. Visualization comparison of depth details. (a)(c) Input images; (b)(d) output images

表 1KITTI Odometry数据集09和10序列的RMSE比较

Table1. RMSE comparison of 09 and 10 sequences in the KITTI Odometry dataset

表 2深度估计模型的TUM评估结果比较

Table2. Comparison of TUM evaluation results for depth estimation model

相关论文

相关资讯

关于本站 Cookie 的使用提示

全站搜索

融合扩张卷积网络与SLAM的无监督单目深度估计下载： 1189次

图 8. 深度预测的定性比较。(a) RGB输入图像;(b) Garg等^[11]的方法;(c) sfmlearner方法^[4];(d)本文方法;(e) ground truth

Fig. 8. Qualitative comparison of depth prediction. (a) RGB input image; (b) method of Garg et al.^[11]; (c) sfmlearner method^[4]; (d) our method; (e) ground truth