Self-Supervised Monocular Scene Decomposition and Depth Estimation
In 3DV 2021


Sadra Safadoust
Fatma Güney


KUIS AI Center, Koç University





Method overview figure

Self-supervised monocular depth estimation approaches either ignore independently moving objects in the scene or need a separate segmentation step to identify them. We propose MonoDepthSeg to jointly estimate depth and segment moving objects from monocular video without using any ground-truth labels. We decompose the scene into a fixed number of components where each component corresponds to a region on the image with its own transformation matrix representing its motion. We estimate both the mask and the motion of each component efficiently with a shared encoder. We evaluate our method on three driving datasets and show that our model clearly improves depth estimation while decomposing the scene into separately moving components.


Input ImageInput
Image
Our Scene Decomposition Our Scene Decomposition
Monodepth2's Depth EstimationMonodepth2's Depth Estimation
Our Depth EstimationOur Depth Estimation

Monocular depth estimation methods assume a static scene by relying on the ego-motion to explain the scene and fail in foreground regions with independently moving objects (bottom-left: Monodepth2). By decomposing the scene into a set of components, we estimate a separate rigid transformation for each component, representing its motion. This improves the results in regions with moving objects (bottom-right) while simultaneously recovering a decomposition of the scene, mostly corresponding to moving regions (top-right).


Input
Image
Our Scene Decomposition
Monodepth2's Depth Estimation
Our Depth Estimation


Download Video Download Button


Paper

Paper Thumbnail

Self-Supervised Monocular Scene Decomposition and Depth Estimation

Sadra Safadoust and Fatma Güney

In 3DV, 2021.

@INPROCEEDINGS{monodepthseg,
  author={Safadoust, Sadra and Güney, Fatma},
  booktitle={International Conference on 3D Vision (3DV)},
  title={Self-Supervised Monocular Scene Decomposition and Depth Estimation},
  year={2021},
  pages={627-636}}
          

Acknowledgements

This project has received funding from KUIS AI Center, TÜBİTAK (118C256), EU Horizon 2020 under Marie Skłodowska-Curie grant (898466).