Publication year and journal Link
Contribution

- Deep learning based SLAM system which consists of recurrent iterative updates of camera pose and pixel-wise depth through a Dense Bundle Adjustment layer.
- It has state-of-the-art performance, outperforming existing SLAM systems, classical or learning-based, on challenging benchmarks with very large margins.
- Substantially fewer catastrophic failures than prior systems.
- Trained only with monocular input, can directly use stereo or RGB-D input to get improved accuracy without any retraining.
- It builds upon RAFT(optical flow) with two key innovations.
- It can be applied to any arbitrary number of frames enabling global refinement of poses and depth essential for minimizing drift for long trajectories and loop closures.
- It apply the differentiable Dense Bundle Adjustment (DBA) layer, which computes a Gauss-Newton update to camera poses and dense per-pixel depth so as to maximize their compatibility with the current estimate of optical flow.
- Like the direct approach, these slam do not require preprocessing steps to detect and match features between the images, instead use the full image, allowing us to leverage a wider range of information than indirect methods with typically only use corners and edges. It also minimizes the reprojection loss.
Models/Algorithms
Details
