Answers are grounded in this paper and cite their sources.
Nighttime pedestrian detection benefits from complementary visible and thermal sensing, yet effective fusion remains difficult because modality reliability varies with illumination and feature scale. We present a lightweight dual-stream detector that combines scale-specific cross-modal fusion (SSCF), illumination-aware weighting (IAW), and modality complementarity regularization (MCR). SSCF learns distinct visible–thermal interactions at multiple feature levels. IAW uses visible-image brightness to adjust scene-level modality weights. MCR adds a squared-cosine penalty to paired, globally pooled P5 features during training, without changing the inference graph. On LLVIP, the complete model achieves a VOC-AP@0.5 of 0.8821 at a confidence threshold of 0.25. This represents improvements of 11.5 and 7.3 percentage points over the RGB-only and early-fusion baselines, respectively. Component analysis identifies MCR as the principal source of the global AP gain, whereas IAW improves the MCR configuration in medium and bright nighttime subsets. Additional training and evaluation on KAIST Night-20 yields a VOC-AP@0.5 (τ = 0.25) of 0.4774, demonstrating the effectiveness of the proposed design on a second nighttime visible–thermal dataset.
No takes yet. Share an insight, caveat, or question.
Xinqi Ke (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: