Infrared and visible images provide complementary cues for low-light image generation, and physics-based methods can better exploit this complementarity. However, existing physics-based methods still rely heavily on paired data. To address these limitations, we propose Multi-View Routing Mixture-of-Experts (MVR-MoE), a framework for physically consistent TeX generation that learns multi-view structural representations. MVR-MoE mines intrinsic structural cues from infrared and visible images to organize them into multi-view representations, while the proposed multi-view hyper connections expand the feature space and preserve both shared evidence and modality-specific details. A hierarchical routing mechanism further regulates feature utilization at both the structural-view level and the expert level, enabling adaptive exploitation of heterogeneous cues. The predicted temperature, emissivity, and texture maps are constrained by the Stefan-Boltzmann law to ensure physically consistent. Extensive experiments on challenging datasets demonstrate that MVR-MoE achieves superior robustness and discriminability while remaining lightweight.
Figures & tables
Figure 1. Illustration of infrared radiation decomposition for TeX generation. The scene is factorized into temperature (T-Map), material emissivity (e-Map), and texture (X-Map), revealing physically meaningful distinctions that are not obvious in raw infrared measurements.
Figure 2. Overview of MVR-MoE. Infrared and visible images are decomposed into high-frequency (HF), low-frequency (LF), and MPC components, which jointly form a multi-view feature set. Through multi-view hyper connections, these representations are jointly encoded into Structural Representations Z1,2,3,4 , facilitating cross-modality interactions. Routing module performs adaptive reweighting and aggregation over Z1,2,3,4 to construct heterogeneous representations, which are subsequently dispatched to specialized experts. The outputs of all experts are finally integrated to generate TeX with physically consistent.
Figure 3. MVHC. Through hyper connections, HF, LF, MPC interact with the F to form four multi-view branches that preserve both shared consensus and view-specific diversity. For each prediction channel, a feature-level routing module first constructs heterogeneous feature representations via adaptive composition of Structural Representations.
Method
Modality
mAP50
mAP75
mAP50:95
YoLov10 ( Wang et al., 2024 )
RGB
90.0
57.3
52.7
YoLov10 ( Wang et al., 2024 )
IR
96.1
74.8
64.7
MMI-Det ( Zeng et al., 2024 )
IR+RGB
98.9
73.5
64.4
GAN-HA ( Lu et al., 2024 )
IR+RGB
94.3
N/A
56.0
RSDet ( Zhao et al., 2024b )
IR+RGB
95.8
70.4
61.3
ICAFusion ( Shen et al., 2024 )
IR+RGB
N/A
N/A
N/A
Table 1. Object detection performance comparison of MVR-MoE and competing methods on the LLVIP dataset.
Method
Modality
mAP50
mAP75
mAP50:95
YoLov10 ( Wang et al., 2024 )
RGB
66.3
27.7
32.7
YoLov10 ( Wang et al., 2024 )
IR
78.2
39.6
42.3
MMI-Det ( Zeng et al., 2024 )
IR+RGB
79.8
35.8
40.5
RSDet ( Zhao et al., 2024b )
IR+RGB
83.9
40.1
43.8
ICAFusion ( Shen et al., 2024 )
IR+RGB
79.2
36.9
41.4
UniRGB-IR ( Yuan et al., 2025 )
IR+RGB
81.4
40.2
44.1
Table 2. Object detection performance comparison of MVR-MoE and competing methods on the FLIR dataset.
Method
Modality
People
Car
Bus
Motorcycle
Lamp
Truck
mAP50
mAP50:95
YoLov10 ( Wang et al., 2024 )
RGB
44.5
68.7
75.4
49.3
50.1
65.8
86.1
59.0
YoLov10 ( Wang et al., 2024 )
IR
53.1
64.6
72.4
41.1
47.7
61.4
84.5
56.7
CFT ( Qingyun and Zhaokui, 2022 )
IR+RGB
52.0
68.2
79.2
49.9
45.2
69.6
89.8
60.7
TarDAL ( Liu et al., 2022a )
IR+RGB
49.8
65.4
69.5
46.6
43.7
61.1
86.0
56.0
MetaFusion ( Zhao et al., 2023c )
IR+RGB
48.4
66.7
70.5
49.1
46.4
59.9
86.7
56.8
CDDFuse ( Zhao et al., 2023a )
IR+RGB
46.1
69.7
74.2
42.2
44.2
62.7
87.0
56.5
Table 3. Quantitative comparison of MVR-MoE and other methods on the M3FD dataset for object detection, including per-class detection performance.
Figure 4. The figure shows visible images, infrared images, saliency-based fusion, task-driven fusion, high-frequency maps, low-frequency maps, monogenic phase congruency maps, and T-Map, e-Map, and X-Map results under low-light conditions.
Table 4. Ablation study of the proposed MVR-MoE framework. Each component, including Mixture-of-Experts (MoE), Multi-View Structure (MS), Multi-View Hyper Connections (MVHC), and Expert Specialization (ES), is progressively introduced to evaluate its contribution. Δ denotes the improvement over the baseline.
Figure 5. Ground truth and generated TeX maps under supervised and self-supervised settings on HADAR.
Method
Time (ms) ↓
FLOPS (G) ↓
Params (M) ↓
TGFuse ( Rao et al., 2023 )
41.17
318.9
137.30
CDDFuse ( Zhao et al., 2023a )
463.8
1402.0
1.186
DDFM ( Zhao et al., 2023b )
280.8k
1336k
552.700
IGNet ( Li et al., 2023 )
47.46
647.6
7.871
MetaFusion ( Zhao et al., 2023c )
85.40
637.9
0.812
MoE-Fusion ( Cao et al., 2023b )
77.79
229.5
83.450
Table 5. Computational Efficiency Comparison of Different Methods ( ↓ indicates lower is better).
Bharti School of Telecommunications Technology and Management, Indian Institute of Technology Delhi, Delhi, India · School of Engineering, Jawaharlal Nehru University, Delhi, India