SLIDE 1
Learning Approaches to Estimate Depth from RGB
Lecture 5
SLIDE 2 What will we learn - Latest Approaches to Depth Estimation based on Machine Learning (DNNs)
- Why do we need new approaches?
- Paper1 -> CNNs for Depth estimation
- Paper2 -> Semantics for Depth Estimation
- Paper3 -> Differentiable Rendering for Depth Estimation
- Paper4 and Paper5 -> Learned Multi-view geometry
SLIDE 3
SLIDE 4 Courtesy figure: Silvio Savarese.
SLIDE 5
SLIDE 6
Meta - What is important in DNN research? Priors Architecture Loss Data
+
SLIDE 7
Eigen et al., “Depth Map Prediction from a Single Image using a Multi-Scale Deep Network”, NeurIPS14
SLIDE 8
Eigen et al.
SLIDE 9
Tatarchenko et al. “What Do Single-view 3D Reconstruction Networks Learn?”, CVPR19
SLIDE 10
Kato et al. - “Neural 3D Mesh Renderer”, CVPR18
SLIDE 11 Rendering
Parameters
+
3D model
SLIDE 12
Kato et al. - “Neural 3D Mesh Renderer”, CVPR18
SLIDE 13
Godard et al. - “Unsupervised Monocular Depth Estimation with Left-Right Consistency”, CVPR17
SLIDE 14
Zhang et al. - LiStereo: Generate Dense Depth Maps from LIDAR and Stereo Imagery, arxiv
SLIDE 15
Question - What are Latest Trends in Learning Depth from RGB?
SLIDE 16
SLIDE 17 What are Latest Trends in Learning Depth from RGB?
- 1. Large amount of data + Powerful parametric function approximators
(DNNs)
- 2. Exploit semantics
- 3. Differentiable Rendering
- 4. Self supervision from stereo
- 5. Sparse supervision from Lidar
