Direct Sparse Odometry

Engel et al. present Direct Sparse Odometry (DSO), a visual odometry system that merges the direct image-formation model of LSD-SLAM with the sparse keyframe windowed optimization of indirect feature-based pipelines. The formulation minimizes a photometric error directly over pixel intensities rather than reprojected feature locations, and jointly optimizes inverse depth, camera intrinsics, and a full photometric calibration covering exposure time, lens vignetting, and the camera response function.

In contrast to dense direct methods such as DTAM, DSO drops the total-variation smoothness prior on depth and instead samples a fixed budget of pixels uniformly across regions of non-zero intensity gradient. The sampling deliberately spans edges, weakly textured walls, and high- gradient interiors, which lets the optimizer extract signal from surfaces that classical keypoint pipelines would discard. Joint windowed bundle adjustment over keyframes refines pose and inverse-depth simultaneously, keeping the operation real-time on a single CPU.

The authors evaluate on three datasets — the TUM-monoVO benchmark, the EuRoC MAV sequences, and an in-house high-frame-rate corpus — totaling several hours of video. DSO outperforms both ORB-SLAM-style indirect baselines and prior direct methods (LSD-SLAM) on tracking accuracy and robustness, particularly under rolling shutter, motion blur, and exposure changes where photometric calibration is decisive.

DSO became a reference baseline for subsequent direct monocular systems and was extended in follow-up work to stereo, RGB-D, and inertial inputs. It is also a recurring comparison point for learned SLAM systems such as DROID-SLAM and DeepV2D, which inherit the dense-photometric residual but replace the hand-engineered front end with a recurrent network. Within the Tesserae demo corpus, DSO sits at the classical end of the SLAM-to-radiance-field continuum: explicit residuals, sparse geometry, real-time on commodity hardware.