Real trajectory data · E3 SIFT-VLAD · MMOTS-WS dataset · DLR Oberpfaffenhofen
E3 · SIFT-VLAD · RGB only · no HSI fusion · MMOTS-WS · 27 iterations · RMSE 12.69 cm → 4.90 cm
01Drift accumulates
Every pose estimate carries a small error. As the rover moves, errors compound — the estimated trajectory diverges further from the truth with each step.
02A place is recognised
The system compares descriptors from the current frame against past frames. A match above threshold becomes a loop closure candidate.
03Geometric verification
The candidate pair undergoes a 3D Gaussian splat overlap check. Only geometrically consistent matches are accepted as confirmed edges.
04Pose graph correction
Confirmed edges are added to the pose graph. Optimisation redistributes accumulated error across the full trajectory — watch the RMSE drop.
Experiment E3 — Key numbers
12.69 cm
RMSE at iteration 1
4.90 cm
Final RMSE — 27 iterations
27
Loop closure iterations
998
Keyframes in final trajectory
8 m
Sequence extent (closed loop)
1,072
Synchronised RGB-HSI pairs