Work

SPARC: A Protocol and Negative Result on the Spatial Allocation of Registration Correction in Misaligned Infrared–Visible Fusion

Year: 2026
Type: article
Abstract: Misaligned infrared–visible fusion is usually handled by dense registration, and a recent line of work instead gates the correction spatially using a “registration necessity” map. We report a negative result for that premise, obtained with SPARC (Spatial Allocation of Registration Correction), a protocol in which both factors of necessity are measurable: the perturbation is analytic, so the per-pixel correction demand is exact, and task importance is rasterised from detection annotations. Across 1049 pairs of the multi-scenario multi-modality (M3FD) benchmark and three training seeds, the premise does not hold under controlled affine misalignment. Dense correction of the registered candidate is the strongest policy (mean average precision at an intersection-over-union of 0.5, mAP50, 0.6922 vs. 0.5887 without correction), and every gate is below it: a zero-parameter displacement threshold reaches 0.6847, a learned task-importance map 0.6732, the factorised geometry × task map 0.6709, and a matched uniform gate 0.6380. Two structural results explain the ordering: multiplying by geometric demand reduces accuracy relative to the task map alone (paired per-image AP50 Δ=+0.0191, 95% CI [+0.0139,+0.0243]), and a proxy map from earlier work is not worse than our learned map when used as a gate (0.6792). With a real cross-modal registration module (median corner error 8.1 px) the correction gain collapses from +0.099 to +0.018, registration error outweighs geometric-factor error by 15×, and gating becomes statistically indistinguishable from dense correction. A calibrated error sweep locates the crossover: gating, even with oracle reliability, only helps once the registration error exceeds ≈12 px, where correction itself is worthless and its gain is at most +0.017. On measured RGB–thermal parallax of the multi-spectral stereo (MS2) dataset (median 25.5 px per frame) the same ordering holds under cross-modal mutual information: a global model applied densely (0.4193) beats its gated application (0.4057), and the per-pixel depth warp is best (0.4399). A controlled mixture of magnitudes on the full evaluation set (a quarter of the pairs below 4 px) shows that the harmful regime is confined to <2 px, that correction already pays at 2–4 px, and that no magnitude-threshold rule beats always-correct, so the decision “correct or not” does not survive as a policy. Separately, the optimal correction strength depends on the global misalignment magnitude, but that dependence does not produce a better policy than correcting densely. We release the protocol, all gate baselines and per-image predictions.
Source: Sensors
Language: English
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7235031372
FWCI:
Cites: 36
Cited by:
Related to: 10