PaperReading

Unsupervised Monocular Depth Learning in Dynamic Scenes

November 2020

tl;dr: Learning depth and motion without explicit semantic masks.

Overall impression

Challenges in SfM: textureless areas, occlusions, reflections, and moving objects. Most previous work deal with dynamic scene by semantic segmentation and ignore pixels in them. Monodepth2 proposed a method to detect pixels do not change across frames to exclude these regions in training.

This paper is the extension of LearnK to dynamic scenes. This paper proposed a sparsity regularization on motion field and it seems to work quite well. It learns four parameters, 2 for optical centers (principal points) and 2 for focal length.

Key ideas

• Motion network
  • Input (DepthNet output + RGB) * 2 = 8 ch
  • Output
    • Egomotion M_ego: 6 DoF, 1x1x6.
    • Object motion translational map $T_{obj}(u, v)$, HxWx3. Very sparse, mostly zero.
    • Total motion field $T(u, v) = T_{obj} (u, v) + T_{ego}$
• Motion reg on T_obj(u, v). Note that this is not on the entire motion field $T(u, v)$.
  • Group smoothness loss: norm of gradients in x and y direction
  • 1/2 sparsity loss: regularization for the residual translation fields. It encourages more sparsity than L1 loss.
• Depth regularization: edge aware smoothness regularization. Regularization is weaker around pixels where color variation is higher.
• Consistency loss: encourages the forward and backward motion between any pair of frames to be opposite of each other

  • $\alpha \frac{

    RR_{inv} - \mathbb{1}

    ^2}{

    R - \mathbb{1}

    ^2 +

    R_{inv} - \mathbb{1}

    ^2} + \beta\frac{

    R_{inv}T + T_{inv}

    ^2}{

    T

    ^2 +

    T_{inv}

    ^2}$

  • Basically we want to penalize (RRinv - 1) and (Rinv T + Tinv), but we need to normalize them.
• RGB loss (photometric loss)
  • Occlusion aware photometric loss
  • SSIM loss

Technical details

• Softplus activation for depth: $z(l) = \log(1+e^l)$
• Cityscape contains the most dynamic scenes. KITTI mostly static scenes.
• Motion reg losses encourage a piecewise constant object motion field. This makes sense if objects are not rotating or if not fast rotating.
• L-1/2 norm approaches L1 norm for small T motion field value, and approaches 1/2 when larger. This is actually very similar to the “Depth guided L1 loss” in KM3D-Net

Notes

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