基于深度学习的特征匹配与位姿估计

• 1. Learning-based feature matching（基于学习的特征匹配）
• • 1.1. Motivation
  • 1.2. Feature matching pipeline
  • 1.3. Why deep learning?
  • 1.4. What is deep learning?
  • 1.5. How to use deep learning to solve feature matching?
  • 1.6. What is a detector？
  • 1.7. What is a good detector
  • 1.8. How to use CNNs
  • 1.9. Uniqueness
  • 1.10. Repeatability
  • 1.11. Performance comparison
  • 1.12. How to use deep learning to solve feature matching?
  • 1.13. What is a descriptor？
  • 1.14. CNN-based descriptors
  • 1.15. Training descriptors
  • 1.16. Learned vs. handcrafted
  • 1.17. Transformation-invariant descriptors
  • 1.18. Application in visual localization
  • 1.19. Where is training data from?
  • 1.20. Camera pose supervision
  • 1.21. How to use deep learning to solve feature matching?
  • 1.22. What is matcher？
  • 1.23. Problem with RANSAC
  • 1.24. Learning-based matcher
• 2. Learning-based object pose estimation（基于学习的物体6DOF位姿估计）
• • 2.1. Problem definition
  • 2.2. Traditional approaches
  • 2.3. Iterative matching
  • 2.4. Feature matching
  • 2.5. Template matching
  • 2.6. Challenges for traditional methods
  • 2.7. Recent advances: deep learning
  • 2.8. Deep learning for object pose estimation
  • 2.9. New approach to old ideas
  • 2.10. Limitations of heatmap representation
  • 2.11. Vector-filed representation of keypoints
  • 2.12. Uncertainty-aware PnP
  • 2.13. Quantitative comparison
  • 2.14. Comparison to direct regression
• 3.0. Summary
• • 3.1. Most 3D vision problems boil down to correspondence problems
  • 3.2. Deep nets are good at learning better representations, resulting in better correspondences.
  • 3.3. Papers & code at
• 4.0. Discussions

1. Learning-based feature matching（基于学习的特征匹配）

1.1. Motivation

1.2. Feature matching pipeline

1.3. Why deep learning?

1.4. What is deep learning?

1.5. How to use deep learning to solve feature matching?

1.6. What is a detector？

1.7. What is a good detector

1.8. How to use CNNs

1.9. Uniqueness

1.10. Repeatability

1.11. Performance comparison

1.12. How to use deep learning to solve feature matching?

1.13. What is a descriptor？

1.14. CNN-based descriptors

1.15. Training descriptors

1.16. Learned vs. handcrafted

1.17. Transformation-invariant descriptors

1.18. Application in visual localization

1.19. Where is training data from?

1.20. Camera pose supervision

1.21. How to use deep learning to solve feature matching?

1.22. What is matcher？

1.23. Problem with RANSAC

1.24. Learning-based matcher

2. Learning-based object pose estimation（基于学习的物体6DOF位姿估计）

2.1. Problem definition

2.2. Traditional approaches

2.3. Iterative matching

2.4. Feature matching

2.5. Template matching

2.6. Challenges for traditional methods

2.7. Recent advances: deep learning

2.8. Deep learning for object pose estimation

2.9. New approach to old ideas

2.10. Limitations of heatmap representation

2.11. Vector-filed representation of keypoints

2.12. Uncertainty-aware PnP

2.13. Quantitative comparison

2.14. Comparison to direct regression

3.0. Summary

3.1. Most 3D vision problems boil down to correspondence problems

3.2. Deep nets are good at learning better representations, resulting in better correspondences.

3.3. Papers & code at

4.0. Discussions

Shall we learn everything in a single network or make some modules in the traditional pipeline learnable.

标签：What,匹配,use,deep,learning,深度,based,位姿,matching
来源： https://blog.csdn.net/qq_60225495/article/details/119426713