ReDiffNet: Differential RGB-Infrared Learning for Low-Light UAV Oriented Vehicle Detection

Qifan Zhang; Ziran Zhou; Ruijie Li; Jincheng Tang; Hao Wang; Qihao Qiao; Chunliu Wang

cs.LGOct 4, 2026

ReDiffNet: Differential RGB-Infrared Learning for Low-Light UAV Oriented Vehicle Detection

Authors: Qifan Zhang, Ziran Zhou, Ruijie Li, Jincheng Tang, Hao Wang, Qihao Qiao, Chunliu Wang

Organizations: Dalian Maritime University, Dalian, China · The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Hubei University of Economics, Wuhan, China

Abstract

Low-light UAV-based RGB-infrared oriented small-vehicle detection is important for nighttime traffic monitoring, emergency response, and urban inspection. Illumination variations, headlight glare, local shadows, and thermal-response degradation cause spatially varying modality reliability, while the small visual extent of vehicles further weakens boundaries, orientation cues, and thermal responses. Accordingly, selecting trustworthy observations based on local modality reliability while further exploiting complementary discriminative information in regions with ambiguous modality preference is key to constructing effective multimodal representations. Based on this insight, we propose ReDiffNet, a reliability-conditioned differential representation network in which modality reliability guides both evidence selection and complementary recovery. Specifically, degradation-aware reliability learning estimates relative spatial reliability, uncertainty-guided differential recovery exploits cross-modal differences to recover complementary cues in ambiguous regions, and reliability-conditioned reconstruction integrates retained and recovered evidence into a unified representation. ReDiffNet achieves 85.3% and 73.9% mAP50 on DroneVehicle and VEDAI, respectively, supporting its effectiveness.

Figures & tables

Figure 1: Illustration of the three core modules in ReDiffNet: (a) DARL for modality reliability estimation, (b) UGDR for complementary cue recovery, and (c) RCR for reliable feature reconstruction.

Method	Pub. + Year	RGB	Infrared	$\mathrm{mAP}_{50}$
RetinaNet [ 10 ]	ICCV 2017	$\checkmark$	$\times$	47.1
R3Det [ 11 ]	AAAI 2021	$\checkmark$	$\times$	60.8
S2ANet [ 12 ]	TGRS 2021	$\checkmark$	$\times$	61.0
Faster R-CNN [ 13 ]	TPAMI 2017	$\checkmark$	$\times$	55.9
RoITransformer [ 14 ]	CVPR 2019	$\checkmark$	$\times$	61.6
Oriented R-CNN [ 15 ]	ICCV 2021	$\checkmark$	$\times$	60.8

Table 1: Comparison with state-of-the-art methods on DroneVehicle. RGB and Infrared indicate the input modalities used by each method. All

\mathrm{mAP}_{50}

values are percentages.

Method	RGB	Infrared	$\mathrm{mAP}_{50}$
RetinaNet [ 10 ]	$\checkmark$	$\times$	20.7
S2ANet [ 12 ]	$\checkmark$	$\times$	44.5
Faster R-CNN [ 13 ]	$\checkmark$	$\times$	61.5
RoITransformer [ 14 ]	$\checkmark$	$\times$	65.4
Oriented R-CNN [ 15 ]	$\checkmark$	$\times$	66.4
RetinaNet [ 10 ]	$\times$	$\checkmark$	18.7

Table 2: Comparison on the VEDAI dataset. All values are

\mathrm{mAP}_{50}

Method	Params (M)	FLOPs (G)	FPS	$\mathrm{mAP}_{50}$
TSFADet [ 5 ]	104.7	109.8	18.6	73.9
CAGTDet [ 22 ]	–	120.6	17.8	74.6
C 2 Former [ 6 ]	100.8	89.9	–	74.2
DMM+S 2 A-Net [ 8 ]	87.97	–	–	79.4
CoDAF [ 20 ]	67.3	224.9	58.1	78.6
ReDiffNet (Ours)	10.53	27.42	106.9	85.3

Table 3: Reported model size and detection performance on DroneVehicle. All

\mathrm{mAP}_{50}

values are percentages.

Method	Precision	Recall	$F_{1}$	$\mathrm{mAP}_{50}$	$\mathrm{mAP}_{50:95}$
DARL removed	80.18	77.95	79.05	81.71	67.30
UGDR removed	79.53	75.95	77.70	80.48	66.16
RCR removed	79.12	76.86	77.98	80.91	66.62
Detection objective only	79.12	78.50	78.80	83.10	68.40
Full model	81.79	80.84	81.31	85.30	70.48

Table 4: Ablation study on DroneVehicle. All variants use the same training and evaluation protocol.

Condition	Factor	$\mathrm{mAP}_{50}$	$\mathrm{mAP}_{50:95}$	Drop 50
Clean	–	85.21	70.19	–
RGB darkening	0.75	84.80	69.80	0.41
RGB darkening	0.50	84.00	68.89	1.21
RGB darkening	0.30	82.77	67.63	2.44
IR contrast reduction	0.80	85.10	70.05	0.12
IR contrast reduction	0.60	84.66	69.62	0.55

Table 5: ReDiffNet under synthetic modality degradation on the DroneVehicle validation set; Drop 50 is in percentage points.

Explore similar work

LER-YOLO: Reliability-Aware Expert Routing for Misaligned RGB-Infrared UAV Detection
May 20, 2026Liming Hou, Yueping Peng, Hexiang Hao +6Drone Detection Unmanned Aerial Vehicle Detection
Engineering University of PAP, Xi’an 710086, China · Unit Command Department, Officers College of PAP, Chengdu 610213, China
EGM-Det: Entropy-Guided Multimodal Adaptive Fusion for UAV RGB-IR Object Detection
Aug 12, 2026Cunzheng Fan, Dawei Yan, Guanlin Wang +4Multimodal Fusion Rgb-Thermal Object Detection
School of Cybersecurity, Northwestern Polytechnical University, Xi’an 710072, China · School of Automation and Software Engineering, Shanxi University, Taiyuan 030006, China
Tri-Modal Fusion Transformers for UAV-based Object Detection
Apr 17, 2026Craig Iaboni, Pramod AbichandaniMulti-Sensor Fusion Drone Detection
New Jersey Institute of Technology

May 20, 2026, cs.CVJ/K move · Enter open · S save

LER-YOLO: Reliability-Aware Expert Routing for Misaligned RGB-Infrared UAV Detection

Liming Hou, Yueping Peng, Hexiang Hao, Ji Wang+5

Engineering University of PAP, Xi’an 710086, China · Unit Command Department, Officers College of PAP, Chengdu 610213, China

Drone Detection Unmanned Aerial Vehicle Detection Infrared Yolo26 Distributional Alignment Experts

Open paper AskPDF arXiv

Detecting small unmanned aerial vehicles from RGB-infrared remote-sensing pairs remains challenging due to tiny target scale, cluttered backgrounds, and spatial misalignment between heterogeneous sensors. Existing bimodal detectors often align or fuse features without assessing the reliability of local cross-sensor correspondence, allowing mismatch artifacts to propagate into the detection head. To address this issue, we propose LER-YOLO, a reliability-aware sparse mixture-of-experts framework for misaligned RGB-infrared UAV detection. LER-YOLO first introduces an Uncertainty-Aware Target Alignment module that resamples visible features toward the infrared reference and estimates a spatial reliability map. This reliability prior is then used by a Reliability-Guided Sparse MoE Fusion module to adaptively select k experts from RGB-dominant, infrared-dominant, and interactive fusion experts, enabling trustworthy cross-modal interaction while suppressing unreliable fusion. Experiments on the public MBU benchmark under a YOLOv5s-family protocol show that LER-YOLO achieves 89.7+/-0.2% AP50 over three independent seeds, with a best result of 89.9%. Extensive ablations, parameter-matched comparisons, synthetic-shift evaluations, and complexity analysis demonstrate that the gains mainly come from reliability-guided expert routing rather than increased model capacity.

Abstract

Explore similar work

LER-YOLO: Reliability-Aware Expert Routing for Misaligned RGB-Infrared UAV Detection

EGM-Det: Entropy-Guided Multimodal Adaptive Fusion for UAV RGB-IR Object Detection

Tri-Modal Fusion Transformers for UAV-based Object Detection