S-PTAM

Pire 2017 · Paper

One-line summary — S-PTAM extends PTAM’s parallel tracking-and-mapping paradigm to stereo cameras, delivering a complete, real-time, ROS-released stereo SLAM system with metric scale from the baseline and appearance-based loop closure.

Problem

Monocular SLAM suffers from the bootstrapping problem (delayed, delicate map initialization), scale ambiguity and scale drift, and fragility under dynamic objects and fast motion. Stereo fixes all of these: triangulating depth from a single stereo view initializes points with low uncertainty and undelayed, and recovers the real scale — “essential for robots which have to interact with their surrounding workspace.” What was missing was a mature, real-time, feature-based stereo SLAM that is robust indoors, outdoors, with dynamic objects, changing light and large loops — and that a robotics lab could actually deploy through ROS. S-PTAM fills that gap.

Method & architecture

Three threads share the map: tracking, local mapping, and loop closing. The global frame is set at the first camera pose; the initial map is triangulated from the first stereo pair (no bootstrapping).

• Features: GFTT corners described with binary BRISK descriptors — chosen from an exhaustive detector/descriptor evaluation on KITTI in the paper; extraction for the two images runs in two parallel threads.
• Tracking (per frame): map points inside the predicted frustum (prediction from wheel odometry or a decaying velocity model) are projected and matched to features by Hamming distance within grid cells (spatial hashing); only points in the covisibility area of nearby keyframes are considered, with frustum culling and a 45° viewing-angle test. The pose is then refined by composing the previous pose with a relative motion $\mu = (t_x, t_y, t_z, \theta_{roll}, \theta_{pitch}, \theta_{yaw})^T$ found by robust Gauss-Newton on the reprojection error: $$\mu' = \operatorname*{argmin}_{\mu} \sum_{i \in S} \rho\left(J_i\, \mu - \Delta z_i(\mu_{\text{prev}})\right),$$ with Huber cost $\rho$, minimized by Levenberg-Marquardt (g2o); $J$ chains the projection derivative with the pose derivative $\partial x_i^C / \partial \mu_j = G_j\, E^{CW}_{\text{prev}}\, x_i^W$.
• Keyframes and immediate point creation: a frame becomes a keyframe when tracked points fall below 90% of those tracked by the last keyframe; unmatched stereo features are triangulated and inserted into the map immediately in the tracking thread — unlike PTAM, which defers this to the mapping thread and can lose tracking while the queue is congested.
• Local mapping: Local Bundle Adjustment refines up to ten queued keyframes plus nearby already-refined keyframes and their visible points, minimizing the same robust reprojection error jointly over poses and points, $\left(\mu'_{j=2\ldots N}, x'^W_{i=1\ldots M}\right) = \operatorname{argmin} \sum_j \sum_{i \in S_j} \rho_{\sigma_T}(\psi_{ji})$, with the first keyframe fixed (it defines the world frame). Bad points are removed; new point-keyframe measurements are actively searched to strengthen the graph.
• Loop closure: keyframes are described as bags of binary words (DBoW2, $\alpha = 0.3$); a candidate must pass the normalized similarity score $\eta(v_i, v_j) = s(v_i, v_j) / s(v_i, v_{i-1})$ where $s(v_i, v_j) = 1 - \frac{1}{2}\left\lVert \frac{v_i}{|v_i|} - \frac{v_j}{|v_j|} \right\rVert$ is the L1 similarity in $[0,1]$. Geometric validation runs P3P RANSAC over 3D-2D correspondences (loop accepted only if 80% inliers), then a full PnP + nonlinear refinement gives the relative transform $T_{C^C C^\ell}$. Correction propagates that transform along the loop with Slerp-interpolated weighting (stronger near the loop point), followed by pose graph optimization (g2o) and map point correction — while a “safe mapping window” lets tracking and LBA keep running, pausing them only briefly for the final map update.

Results

Experiments on KITTI (00–10) and the Indoor Level 7 S-Block dataset, on a laptop with an i7 @ 2.8 GHz and 16 GB RAM:

• KITTI benchmark: translation error 1.19% vs. ORB-SLAM2 (stereo) 1.15% and S-LSD-SLAM 1.20%; rotation error 0.0025 deg/m — the best of the three (ORB-SLAM2 0.0027, S-LSD-SLAM 0.0033).
• Loop closure: geometric validation rejected false positives with 100% precision on all loop sequences (00, 02, 05, 06, 07, 09); over the ~4 km sequence 00 trajectory the maximum absolute localization error stayed below 15 m.
• Runtime: the tracking thread runs at ~18 Hz including loop-closure overhead.
• Level 7 S-Block (wheeled robot, office loops over 30+ minutes): accuracy comparable to ORB-SLAM2, with similar error peaks in the same areas.

Why it matters for SLAM

S-PTAM is a bridge between the PTAM era and modern stereo SLAM: it showed that the tracking/mapping decomposition scales cleanly to stereo and large outdoor trajectories, and it was one of the first complete open-source stereo SLAM systems with loop closure that a robotics lab could deploy through ROS. It served as a common ground-robot baseline before ORB-SLAM2 became the dominant feature-based reference, and details like immediate point creation in the tracking thread and the safe-window loop correction remain instructive engineering lessons.

Related

• PTAM
• ORB-SLAM2
• StereoMSCKF
• Disparity vs Depth
• Stereo rectification
• Visual Place Recognition (VPR)