[Zador-Pataki]

Motivation

Use FlexGlue for sparse keyframe queries and RoMaV2 refinement for accurate tracking and loop-closure matches. Nearby video frames typically require less reasoning to match, so we exploit FlexGlue’s configurable depth to use fewer transformer layers for sequential matching while retaining the full model for loop closures. Early rejection saves runtime by rejecting unsuitable pairs before matching finishes, while pair rejection protects against visual aliasing.

Changes

• Make FlexGlue the default; retain RoMaV2 through matcher.name.
• Use 6 FlexGlue transformer layers for keyframing and temporal tracking, and all 9 for retrieval-based loop-closure matching.
• Reject pairs whose FlexGlue pair-confidence score is below 0.8. An early-stop decay of 0.2 controls how the rejection threshold changes across layers. Adjacent keyframe pairs are always kept to preserve temporal connectivity.
• Reuse image features and the RoMa refinement implementation, with compiled inference and asynchronous image transfers.

Known regression and follow-up

In some scenarios that are impossible to match, RoMaV2 hallucinates matches while FlexGlue produces very few noisy matches. FlexGlue's pair-rejection head flags these cases, but adjacent keyframe pairs are currently protected from rejection.

FlexGlue largely ignores the unmatchable frame, while local loop closures constrain the frames before and after it. This often gives better trajectory accuracy than RoMaV2, but the unsupported frame can float away and contribute an evaluation outlier. We could use the doppelganger head to reject the pair, then filter the unsupported frame or constrain it with a temporal smoothness prior.

Runtime

RTX 4090, three 129-frame ETH3D clips; two measured repeats after warm-up. Both use the same tracking timeline and LC pairs. Coarse inputs: FlexGlue 512 px, RoMa 560 px; refinement: 1200 px.

Stage	Unit	RoMaV2	FlexGlue	Speedup
Keyframe selection	ms/input-frame transition	27.44	23.93	1.15×
Track propagation	ms/new keyframe	251.44	209.76	1.20×
Loop-closure matching	ms/candidate pair	69.83	57.84	1.21×

Benchmark results: RoMaV2 and FlexGlue

Both methods use LC-only feature retention and deduplication, with loop-closure weight 1×.

ETH3D monocular · 55 scenes · Pose AUC

Method	ATE (m) ↓	5 cm	10 cm	50 cm	1 m	10 m
RoMaV2	0.268	80.32	87.85	95.86	97.35	99.28
FlexGlue	0.066	81.64	87.78	94.79	96.79	99.54

EuRoC monocular · 11 scenes · Pose AUC

Method	ATE (m) ↓	5 cm	10 cm	50 cm	1 m	10 m
RoMaV2	0.044	28.44	59.74	91.91	95.95	99.60
FlexGlue	0.044	29.32	60.40	92.04	96.02	99.60

CroCoDL · 64 scenes · W-AUC at 5%

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	0.233	95.72	97.27	97.39	96.56	98.21
FlexGlue	0.284	95.38	96.75	96.70	96.60	98.03

LaMAR 300 · 50 scenes · W-AUC at 5%

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	0.878	92.43	93.08	91.79	91.03	90.41
FlexGlue	0.916	92.78	93.87	92.90	91.99	90.83

LaMAR older selection · 53 evaluated trajectories · W-AUC at 5%

These results do not cover the current complete testset (128 recordings).

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	1.452	92.42	92.83	92.38	92.59	89.76
FlexGlue	1.361	92.51	93.18	93.08	92.91	90.88

Spires rectified monocular · 10 scenes · W-AUC at 5%

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	0.130	98.98	99.10	99.11	99.10	99.54
FlexGlue	0.123	98.98	99.08	99.06	99.06	99.52

Hilti forward monocular · 30 scenes · W-AUC at 5%

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	0.822	93.96	93.96	93.29	92.44	94.46
FlexGlue	0.518	94.70	95.41	95.63	95.01	96.36

Hilti three-view · 30 scenes · W-AUC at 5%

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	0.206	94.80	96.22	96.94	97.21	98.13
FlexGlue	0.253	94.55	95.85	96.56	96.81	97.93

Spires three-camera rig · 13 scenes · W-AUC at 5%

Method	ATE (m) ↓	10 m	25 m	50 m	100 m	Full
RoMaV2	0.119	99.15	99.24	99.26	99.28	99.59
FlexGlue	0.088	99.17	99.30	99.36	99.43	99.68