[UnDeepVO]_Monocular_Visual_Odometry_through_Unsupervised_Deep_Learning.md

UnDeepVO: Monocular Visual Odometry through Unsupervised Deep Learning

arXiv

System overview

空间重建+时间重建

Loss

1. Spatial Image Losses of a Stereo Image Pair
   1. 空间重建可以产生绝对尺度，recover scaled depth，solve scaling ambiguity issue
   2. Photometric Consistency Loss 左右图像依靠视差互相重建，计算相似度 SSIM
   3. Disparity Consistency Loss 计算是视差图的相似度
   4. Pose Consistency Loss 左右图像都预测pose，两种pose应一致
2. Temporal Image Losses of Consecutive Monocular Images
   1. Photometric Consistency Loss $$p_{k+1}=K(^{k+1}T_k)DK^{-1}p_k$$ 其中k为时间，K为内参数，T为transform矩阵，D为深度
   2. 3D Geometric Registration Loss 在k下的三维点可转移到k+1下，反之亦然