| PoseNet |
GoogleNet+FC，回归位姿 |
$loss(I)={| x-\hat{x}|}_2 + {| q-\hat{q}|}_2$ |
Pose |
7 Scenes |
0.44m |
5.24° |
|
|
— |
— |
— |
— |
— |
— |
|
ICCV15 |
位姿回归 |
| |
|
|
|
Cambridge |
2.4m |
3.38° |
|
|
— |
— |
— |
— |
— |
— |
|
|
|
| CNN-SLAM |
ResNet50预测深度图 |
Reverse Huber Loss |
Depth |
ICL-NUIL |
— |
— |
0.202 |
|
|
|
|
|
|
76.96% (a <1.1) |
|
CVPR17 |
|
| |
|
|
|
TUM RGB-D SLAM |
— |
— |
0.333 |
|
|
|
|
|
|
78.68% (a < 1.1) |
|
|
|
| LIFT |
特征提取网络 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ECCV16 |
|
| MagicPoint |
特征提取网络 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
arXiv17 |
|
| GCN |
特征提取网络 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ICRA18 |
|
| DeMoN |
en-de链式结构 |
$loss=|s\xi-\hat{\xi}|+|n-\hat{n}|+|w-\hat{w}|$ |
Depth, normal, optical flow |
MVS, Scenes11, TUM RGB-D, Sun3D, NYUv2 |
— |
|
|
|
|
|
|
|
|
|
|
CVPR17 |
|
| SfMLearner |
DepthCNN+PoseCNN |
光度误差+深度smooth |
Unsupervised |
KITTI Odom |
— |
— |
0.021(seq09) 0.020(seq10) |
|
— |
— |
— |
— |
— |
— |
|
CVPR17 |
|
| |
|
|
|
KITTI Eigen |
— |
— |
— |
— |
— |
— |
— |
0.208 |
6.856 |
67.8% |
|
|
|
| |
|
|
|
Make3D(泛化) |
— |
— |
— |
— |
— |
— |
— |
0.383 |
10.47 |
— |
|
|
|
| SfM-Net |
显式预测运动目标位姿，单/双目训练 |
光度误差+深度smooth+前后一致性约束 |
Unsupervised, depth, pose, optical flow, object motion |
TUM RGB-D SLAM |
|
|
|
|
|
|
|
|
|
|
|
arXiv17 |
缺乏定量实验 |
| Depth-Feat-VO |
双目训练，DepthCNN+PoseCNN |
光度误差+特征重构误差+深度smooth |
Unsupervised |
KITTI Odom |
— |
— |
— |
— |
11.93%(seq09) 12.45%(seq10) |
— |
3.91(seq 09), 3.46(seq 10) |
|
|
|
|
CVPR18 |
|
| |
|
|
|
KITTI Eigen |
— |
— |
— |
— |
— |
— |
— |
0.144 |
5.869 |
80.3% |
|
|
|
| Geo-Net |
预测光流为主，刚体光流+非刚体光流预测 |
光度误差(L1+SSIM)+深度smooth+几何一致性约束 |
Unsupervised |
KITTI Odom |
— |
— |
0.012(seq09) 0.012(seq10) |
— |
— |
— |
— |
— |
— |
— |
|
CVPR18 |
|
| |
|
|
|
KITTI Eigen |
|
|
|
|
|
|
|
0.155 |
5.857 |
79.3% |
|
|
|
| BA-Net |
feature-metric |
feature-metric+位姿L2监督+深度Huber监督 |
Pose, Depth |
KITTI Odom |
— |
— |
0.019(km) |
— |
— |
— |
— |
— |
— |
— |
|
ICLR19 |
|
| |
|
|
|
KITTI Eigen |
— |
— |
— |
— |
— |
— |
— |
0.083 |
3.640 |
— |
|
|
|
| |
|
|
|
ScanNet |
3.39(cm) |
1.018° |
— |
— |
— |
— |
— |
0.161 |
0.346 |
— |
|
|
|
| Code-SLAM |
VAE深度估计 |
光度误差+几何残差 |
Unsupervised |
|
|
|
|
|
|
|
|
|
|
|
|
CVPR18 |
|
| Left-Right consistency |
双目训练 |
光度误差(L1+SSIM)+视差smooth+视差一致性约束 |
Unsupervised |
KITTI Stereo |
— |
— |
— |
— |
— |
— |
— |
0.068 |
4.392 |
87.9% |
|
CVPR17 |
|
| |
|
|
|
KITTI Eigen |
— |
— |
— |
— |
— |
— |
— |
0.148 |
4.935 |
86.1% |
|
|
|
| UnDeepVO |
双目训练，预测视差，位姿旋转平移分离 |
左右目光度误差+前后帧光度误差+视差一致性约束+位姿一致性约束 |
Unsupervised |
KITTI Odom |
— |
— |
— |
— |
4.07% |
— |
2.02 |
— |
— |
— |
|
ICRA18 |
|
| |
|
|
|
KITTI Eigen |
— |
— |
— |
— |
— |
— |
— |
0.183 |
6.57 |
— |
|
|
|
| Depth From Wild |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
arXiv19 |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|