(plan) The Deep SLAM Era: What Should Be Learnt?
1 min readJun 12, 2020
NOTE
- Contents were selected first, I’m under making summaries of them, published later.
- The below contents mainly consider ego-motion estimation or metric localization, not the place recognition.
- For place recognition, the deep learning-oriented scheme has been known to be a good choice because it effectively summarizes raw and complex sensor data (e.g., NetVLAD).
1. Early works: try to end-to-end regression and make it better
- 15 CVPR PoseNet
- 17 ICRA DeepVO, 18 ICRA UnDeepVO
- 17 CVPR (PoseNet v2) Geometric loss functions for camera pose regression with deep learning
- 18 RAL VLocNet++: Deep Multitask Learning for Semantic Visual Localization and Odometry
- 18 CVPR (MapNet) Geometry-aware learning of maps for camera localization
2. Transition: not direct regress …
- 18 ECCV Relocnet: Continuous metric learning relocalisation using neural nets
3. Lesson: No more deep-only dream
- 19 CVPR Understanding the Limitations of CNN-based Absolute Camera Pose Regression
4. Current Trend: combine the traditional well-defined geometrical methods and SLAM pipeline
- 17 CVPR CNN-SLAM: Real-time dense monocular SLAM with learned depth prediction
- 18 ECCV Deep Virtual Stereo Odometry: Leveraging Deep Depth Prediction for Monocular Direct Sparse Odometry
- 20 ICRA Visual Odometry Revisited: What Should Be Learnt?
- 20 ICRA To Learn or Not to Learn: Visual Localization from Essential Matrices
- 20 RAL (ICRA) GN-Net: The Gauss-Newton Loss for Multi-Weather Relocalization
- 20 RAL (ICRA) DeepFactors: Real-Time Probabilistic Dense Monocular SLAM
- 20 CVPR D3VO: Deep Depth, Deep Pose and Deep Uncertainty for Monocular Visual Odometry