Skip to content

Latest commit

 

History

History
37 lines (25 loc) · 2.45 KB

d3vo.md

File metadata and controls

37 lines (25 loc) · 2.45 KB
Raw

D3VO: Deep Depth, Deep Pose and Deep Uncertainty for Monocular Visual Odometry

May 2020

tl;dr: Add depth estimation and brightness transformation to Monodepth2. The paper incorporate the deep predicted pose into regularization into a backend.

Overall impression

Monocular VO suffers from scale drift and low robustness. The pose from PoseNet in sfm learner and Monodepth2 are robust, but they are not as accurate as geometry based methods. This paper explores on how to combine geometric approach to deep learning approach (aka "hybrid").

Hybrid methods combines deep learning with geometry based methods:

  • D3VO uses the pose and depth prediction in the backend of DSO
  • DF-VO uses both optical flow and depth inside a robust essential matrix estimation loop.
  • KP3D also uses DSO as backend.

As repeated demonstrated as with other hybrid methods, D3VO beats all other end to end methods by a large margin.

VO lacks robustness for low texture area and fast movement. VIO is more robust, but IMUs cannot deliver the metric scale in constant velocity.

Both KP3D and D3VO uses DSO as backned, and KP3D reaches on par performance with DVSO, while D3VO beats DVSO and even achieves comparable to stereo/lidar methods on KITTI odometry.

Key ideas

  • Predicts a photometric uncertainty map to capture regions that may violate the brightness constancy assumption (non-lambertian surface, moving objets). This is related to the flow consistency map in DF-VO and explainability/occlusion mask in SfM-learner.
  • The paper also predicts a brightness transformation parameter (linear scaling) which is critical for some dataset (KITTI is largely stable).
  • D3VO backend is based on DSO.
    • The virtual stereo term optimizes the estimated depth from VO to be consistent with the depth predicted by the proposed deep network.
    • The predicted pose is integrated in a similar fashion to with IMU with a Gaussian model, as a regularizer in the energy function.

Technical details

  • Uses both left/right stereo consistency and temporal consistency. Quaduplets ($I_t$, $I_t^s$, $I_{t-1}$, $I_{t+1}$)
  • Multiscale loss
  • 40k training quadruplets.
  • Uncertainty boosted performance on KITTI the most.

Notes

  • Questions and notes on how to improve/revise the current work