Learning-Based Feature Detection and Matching
Learn to detect features Learn to correspond features between frames via their descriptors
Deep Visual Odometry
Monocular Depth Estimation End-to-End Visual Odometry (CNN-LSTM architecture capable of predicting pose from image sequences)
Learning-Based Calibration
Predict camera intrinsics from a single image, collects features prone to distortion and FOV (outputs intrinsics).
- trains on synthetic data with known intrinsics
- real data with already estimated intrinsics from SfM
- useful for autocalibration Extrinsic Calibration (predict extrinsics of camera in lidar frame)
Loop Closure Detection
Bag of words Place recognition
End-to-End Learning for SLAM
Take in an image sequence and predict depth and pose for each frame.
- Build a map from those depths and poses which can be differentiable Learn a map representation through pose net and a branched off NeRF model to do neural rendering
Outlier Rejection
Learned RANSAC