ICCV WS 2017 (Conference Paper)

A new framework for integrating geometrical measurements of monocular visual SLAM and depth prediction using convolutional neural network (CNN).

• Introduced mesh deformation approach to fuse geometrical measurements and depth prediction.
• Reduced the reconstruction error by 38% compared to using CNN-based depth map prediction alone.

Tomoyuki Mukasa, Jiu Xu, and Bjorn Stenger: 

3D Scene Mesh From CNN Depth Predictions And Sparse Monocular SLAM, 
The 2nd Workshop on Geometry Meets Deep Learning, ICCV 2017.