IPN-LightGlue: Illumination-Aware Deep Feature Matching via Adaptive Matchability Modulation

Under adverse illumination, matchability scores may distinguish valid key points from ambiguous candidates less reliably. Fixed decision thresholds can then retain unreliable correspondences or discard useful candidates, destabilizing geometric estimation with random sample consensus. IPN-LightGlue is introduced to address this problem by coupling an illumination perception network (IPN) with adaptive matchability modulation. The IPN uses a frozen MobileNetV3-Small backbone to predict a normalized image-level illumination score, which is converted into a non-positive bias on LightGlue’s matchability logits. This design leaves the feature extractor unchanged and adds 74 000 trainable parameters. On the evaluation hardware, an isolated IPN forward pass takes 3.399 ± 0.829 ms per image, and the complete pipeline incurs a runtime overhead of 4.169% to 8.027% relative to LightGlue. Mean matching accuracy at a 3-pixel threshold (MMA@3px) increases from 79.9% to 85.2% on the HPatches illumination subset. For the multi-illumination data set, which isolates illumination-direction changes from viewpoint variation, MMA@3px rises from 85.2% to 89.5%, whereas mean corner error decreases by 16.4%. The method also reduces mean corner error from 59.75 pixels to 7.14 pixels on the HPatches viewpoint subset, despite the IPN being supervised only on radiometric variation. Neither contrast-limited adaptive histogram equalization nor zero-reference deep curve estimation, applied before matching, improves MMA@3px over LightGlue in these evaluations. The findings support illumination-aware matchability modulation as a way to improve correspondence accuracy and geometric stability without modifying the input image.