Chapter 8: Direct Approaches to Visual SLAM

Shortcomings of Classical Approaches

suboptimal: In the selection of feature points much potentially valuable information contained in the colors of each image is discarded.

They invariably lack robustness: Errors in the point correspondence may have devastating effects on the estimated camera motion

They do not address the highly coupled problems of motion estimation and dense structure estimation.

advantages of direct approaches

more robust to noise because they exploit all available input info

semi-dense geometric reconstruction

typically faster

Compare and contrast direct methods with classical feature-based methods for visual SLAM.

Direct methods use raw pixel intensities for estimation, while feature-based methods extract and match sparse features. Direct methods can provide denser reconstructions and are often more robust, but may struggle with large motions. Feature-based methods are more suited to wide baseline matching but discard much of the image information.

Explain the concept of photometric calibration in the context of Direct Sparse Odometry.

Photometric calibration in DSO accounts for non-linear effects in image formation, including vignetting, gamma correction, and exposure time. It converts measured brightness to linear irradiance, allowing for more accurate intensity comparisons across frames and improving the robustness of direct methods.

Describe the Sim(3) transformation group and why it's useful in monocular SLAM

Sim(3) is the group of 3D similarity transformations, including rotation, translation, and scaling. It's useful in monocular SLAM because it allows explicit handling of scale changes, which are inherent in monocular systems where absolute scale cannot be determined.

How does windowed optimization in Direct Sparse Odometry differ from full bundle adjustment?

Windowed optimization in DSO optimizes only recent frames and points, marginalizing out older information. This allows for real-time performance on long sequences, unlike full bundle adjustment which optimizes all parameters. However, it may lose some global consistency compared to full BA.

Explain the trade-offs between dense, semi-dense, and sparse reconstructions in direct SLAM methods.

Dense reconstructions provide the most complete scene representation but are computationally expensive. Semi-dense methods (like LSD-SLAM) offer a balance, reconstructing areas with sufficient texture. Sparse methods (like DSO) are the fastest and allow for efficient optimization but provide less scene detail.