- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes RouteFuse for infrared–visible fusion under rain, haze, and snow. Infrared evidence is injected as gated residuals relative to a primary multimodal representation. The injection is routed separately in low- and high-frequency bands at five feature scales, with gates conditioned on a learned visible-reliability map, a cross-modal feature relation descriptor, and the diffusion timestep. The predictor is trained with v-prediction plus image-space losses toward a hand-crafted fusion target built from clean visible and infrared images, and is applied once at t=299 at inference. Experiments on AWMM-100K report the best overall scores on six fusion metrics, the best detection results, and the lowest FLOPs and runtime. A mechanistic analysis, including an energy-matched Anti-Reliability vs. Random reallocation, supports the claim that reliability-conditioned allocation matters.
Critical strengths
-
The framing of adverse-weather fusion as reliability-conditioned evidence allocation is clear and well motivated.
-
The energy-matched reallocation experiment (Fig. 3c) is a thoughtful causal test that goes beyond correlation and is rarely seen in fusion papers.
-
Consistent overall gains, particularly on SCD and VIF, together with strong efficiency in FLOPs and runtime.
-
The paper is well written, and the appendix gives detailed loss definitions and weights.
Critical weaknesses
- The “one-step diffusion” framing. (1) At t=299 with T=300 and a cosine schedule, α_t ≈ 0, so x̂₀ ≈ −v̂ and the input x_t is essentially pure noise. The model therefore behaves as a noise-conditioned regressor; (2) A8 shows a gain over a deterministic encoder–decoder, but this could stem from multi-timestep training acting as augmentation or regularization rather than from diffusion modeling per se.
- Weak reliability estimates and limited contribution of adaptive gating. Despite direct supervision toward R*, predicted reliability correlates only weakly with local fidelity (ρ = 0.203 and 0.344), and Fig. 3(a) shows many images with negative correlation. In Table 4, trained static gates (A5) are only marginally worse than the full model (e.g., QMI 0.4460 vs. 0.4529), which suggests that most of the gain comes from the residual pathway (A4) rather than from adaptive reliability-conditioned gating.
- Synthetic-only evaluation. The reliability target requires V_gt, and all test results are on synthetic degradations. Results on real adverse-weather captures (e.g., the real-world portion of AWMM-100K reported by AMG-Fuse) would substantially strengthen the claims.
- Baseline fairness and missing baselines. (1) Six of nine baselines are run from author checkpoints after WeatherRemover preprocessing, without retraining on AWMM-100K, while RouteFuse is trained in-domain; (2) Methods discussed in the related work and designed for this setting are not compared: AWFusion (which introduced AWMM-100K), AWM-Fuse, ControlFusion (which covers rain and haze), TG-ECNet, and MdaIF; and (3) The omission of AWFusion and AWM-Fuse is particularly notable given that they target this exact benchmark.
- Selective reporting. (1) Under haze, RouteFuse trails CAWM-Mamba on QMI, QG, and QAB/F, but Table 1(b) shows only QMI, SSIM, and SCD; (2) On clean M3FD, RouteFuse’s VIF (0.4087) ranks near the bottom, and this is not discussed; and (3) The model has the largest parameter count (213M), which is not mentioned alongside the efficiency claims.
- Evaluation references are unspecified. All six metrics require source images as references, but the paper does not state whether the visible reference is the degraded V_d or the clean V_gt. With V_d, higher scores may reward preserving weather artifacts. With V_gt, the evaluation closely matches the training target x₀, which is constructed from V_gt, an advantage the WR-based pipelines do not share. This must be clarified for Table 1 to be interpretable.
- Minor. (1) Please clarify on which images the YOLO11x detector was trained; (2) DiffFuseNet (Wang et al., 2026, Pattern Recognition 180:114259) could not be located; please provide a link; and (3) The WeatherRemover volume and page metadata appear inconsistent with the available version.
Questions for authors
- Which visible image (V_d or V_gt) is used as the reference for each metric?
- Can you retrain or fine-tune the main baselines on AWMM-100K, and add AWFusion and AWM-Fuse?
- How does RouteFuse compare with the same network trained deterministically under the same multi-noise-level augmentation?
- Can you report results on real adverse-weather images?
- Given the modest gap between A5 and the full model, what evidence shows that input-adaptive reliability conditioning, rather than the residual pathway itself, drives the improvement?