The End Of Knowledge - Vault 6/58 - CVPR - 2020

graph LR classDef main fill:#f9d4f9, font-weight:bold, font-size:14px classDef slam fill:#f9d4d4, font-weight:bold, font-size:14px classDef deep fill:#d4f9d4, font-weight:bold, font-size:14px classDef integration fill:#d4d4f9, font-weight:bold, font-size:14px classDef performance fill:#f9f9d4, font-weight:bold, font-size:14px classDef future fill:#d4f9f9, font-weight:bold, font-size:14px Main[Deep Direct Visual
SLAM] --> A[Visual SLAM
Basics] Main --> B[Deep Learning
Integration] Main --> C[Enhanced SLAM
Methods] Main --> D[Performance and
Evaluation] Main --> E[Future Directions] A --> A1[Direct visual
SLAM reconstructs 3D,
camera motion 1] A --> A2[Classical vs direct
methods: geometric vs
photometric error 2] A --> A3[LSD SLAM:
alternates tracking and
depth estimation 3] A --> A4[Brightness consistency:
minimize pixel intensity
differences 4] A --> A5[Real-time performance
on single CPU
core 5] A --> A6[Large-scale reconstruction
with low drift 6] B --> B1[Deep learning
enhances direct SLAM
methods 8] B --> B2[Single-image depth
prediction improves SLAM 10] B --> B3[Deep learning
recovers absolute scale 11] B --> B4[Pose prediction
aids in tracking 13] B --> B5[Brightness correction
predicts affine transformations 14] B --> B6[Aleatoric uncertainty
downweights unreliable areas 15] C --> C1[Deep Virtual
Stereo Odometry integrates
neural predictions 12] C --> C2[Non-linear factor
graph integrates deep
learning predictions 16] C --> C3[Feature space
transformation robust to
appearance changes 20] C --> C4[Gauss-Newton Net
produces optimization-suited features 21] C --> C5[Multi-weather localization
despite environmental changes 22] C --> C6[Anisotropic uncertainty
in feature matching 28] D --> D1[Drift quantification
using loop closure 7] D --> D2[Trained networks
generalize to new
environments 17] D --> D3[Deep learning-enhanced
monocular outperforms classical
stereo 18] D --> D4[Benchmark datasets
for multi-weather localization 23] D --> D5[Generalization to
unseen weather conditions 24] D --> D6[Simulated performance
transfers to real-world 27] E --> E1[Semantic mapping
labels 3D reconstructions 9] E --> E2[Relocalization challenge:
different weather and
lighting 19] E --> E3[High-precision localization
in various conditions 25] E --> E4[Real-time 3D
mapping for autonomous
systems 26] E --> E5[Basin of
attraction ensures convergence 29] E --> E6[Robustness to
occlusions and environmental
changes 30] class Main main class A,A1,A2,A3,A4,A5,A6 slam class B,B1,B2,B3,B4,B5,B6 deep class C,C1,C2,C3,C4,C5,C6 integration class D,D1,D2,D3,D4,D5,D6 performance class E,E1,E2,E3,E4,E5,E6 future

Resume:

1.- Direct visual SLAM: Uses raw image intensities to reconstruct 3D structure and camera motion, avoiding intermediate point extraction and matching steps.

2.- Classical vs. direct methods: Classical methods use geometric reprojection error, while direct methods minimize photometric color consistency error.

3.- LSD SLAM: Large-scale direct SLAM method that alternates between camera tracking and depth map estimation for keyframes.

4.- Brightness consistency: Direct methods optimize camera motion by minimizing differences in pixel intensities between aligned images.

5.- Real-time performance: LSD SLAM runs on a single CPU core, allowing other cores for depth estimation and optimization.

6.- Large-scale reconstruction: The method can reconstruct large outdoor environments with relatively low drift.

7.- Drift quantification: Sequences looping back to start point allow measurement of total drift in translation, rotation, and scale.

8.- Deep learning integration: Neural networks can enhance direct SLAM methods by predicting depth, pose, and uncertainty.

9.- Semantic mapping: Deep networks can label 3D reconstructions with semantic information like drivable areas, cars, and pedestrians.

10.- Single-image depth prediction: Neural networks can estimate depth from a single image, improving initialization in SLAM.

11.- Scale estimation: Deep learning allows monocular systems to recover absolute scale, previously impossible with single cameras.

12.- Deep Virtual Stereo Odometry: Integrates deep learning depth and pose predictions into classical SLAM pipeline.

13.- Pose prediction: Neural networks can estimate relative pose between consecutive frames, aiding in tracking.

14.- Brightness correction: Networks can predict affine transformations to correct for brightness changes between frames.

15.- Aleatoric uncertainty: Networks can estimate uncertainty in predictions, allowing down-weighting of unreliable areas.

16.- Non-linear factor graph: Integrates deep learning predictions into front-end tracking and back-end optimization of SLAM.

17.- Generalization: Trained networks can generalize to new environments and datasets not seen during training.

18.- Monocular vs. stereo performance: Deep learning-enhanced monocular methods can outperform classical stereo methods.

19.- Relocalization challenge: Recognizing the same location under different weather and lighting conditions is difficult.

20.- Feature space transformation: Networks can transform images into consistent feature spaces robust to appearance changes.

21.- Gauss-Newton Net: Designed to produce features optimally suited for subsequent optimization in Gauss-Newton SLAM algorithms.

22.- Multi-weather localization: Ability to localize in previously mapped environments despite significant changes in lighting and weather.

23.- Benchmark datasets: Created multi-weather localization benchmarks using simulator and real-world data for evaluation.

24.- Generalization to unseen conditions: Methods trained on certain weather conditions can generalize to unseen weather types.

25.- High-precision localization: Achieved unprecedented precision in localizing autonomous systems or cars in various conditions.

26.- Real-time 3D mapping: Enables creation of large-scale, high-resolution 3D maps in real-time for autonomous systems.

27.- Simulated vs. real data: Performance on simulated environments can transfer well to real-world scenarios.

28.- Anisotropic uncertainty: Gauss-Newton loss allows for direction-dependent uncertainties in feature matching.

29.- Basin of attraction: Design of loss functions to ensure convergence even with poor initialization.

30.- Robustness to occlusions: Methods can handle temporary occlusions and changes in the environment between mapping and localization.

Knowledge Vault built byDavid Vivancos 2024