acceptodds
Withdrawn from ICLR 2027

DOME: Distortion-Driven Domain Mixture Experts for All-Weather Multimodal 3D Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable 3D perception under adverse weather is essential for autonomous driving, yet remains challenging due to weather-induced sensor degradation. Different adverse weather conditions introduce heterogeneous physical distortions into LiDAR observations, leading to latent domain shifts that are difficult to characterize with discrete weather categories. Existing weather-aware approaches typically rely on image-based weather classifiers to route features according to predefined weather labels, resulting in a mismatch between semantic weather categories and actual sensor degradation patterns, especially under degraded visual conditions. To address this challenge, we propose DOME, a Distortion-Driven Domain Mixture-of-Experts framework for all-weather multimodal 3D object detection. DOME models LiDAR–4D Radar fusion through latent physical-distortion domain decomposition, where experts are specialized according to distortion patterns rather than predefined weather identities. Specifically, the Physical Distortion-guided Router (PDR) extracts distortion fingerprints from multimodal features and guides Distortion-Specialized Experts (DSE) to perform distortion-aware feature adaptation. Furthermore, we introduce DOME-L with LoRA-based Distortion Experts (LDE), reducing the parameter overhead of distortion-driven MoE while preserving domain-specific adaptation capability. Extensive experiments on K-Radar and OmniHD-Scenes demonstrate strong performance. On K-Radar, DOME improves L4DR by 8.1 points overall, while a capacity-controlled comparison isolates a 1.9-point gain from distortion-aware routing. Additional evaluations demonstrate robustness under continuously varying and held-out degradation conditions, as well as transferability to a different multimodal detection backbone. The open-source code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Autonomous driving, 3D object detection, Robust perception

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_F2ekOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes DOME, a Distortion-Driven Domain Mixture-of-Experts framework for robust LiDAR–4D Radar 3D object detection under adverse weather. Rather than relying on predefined weather categories or image-based weather recognition, DOME learns latent, distortion-aware routing cues directly from multimodal features. Its Physical Distortion-guided Router (PDR) combines global multimodal context with local LiDAR–Radar discrepancies to build a distortion fingerprint, which it then uses to activate Distortion-Specialized Experts through Top-K routing. The paper also introduces DOME-L, a parameter-efficient variant that replaces full experts with LoRA-based distortion experts. Experiments on K-Radar and OmniHD-Scenes show improved all-weather detection performance, with additional analyses on generic-MoE baselines, weather-based routing, continuous and held-out degradation conditions, expert routing behavior, and cross-backbone transfer.

Critical strengths

The main strengths of the paper are that (i) it identifies a meaningful limitation of existing weather-aware adaptation and proposes a practically relevant, camera-free alternative, and (ii) the design of the proposed router is well motivated and examined through targeted analyses beyond aggregate accuracy.

First, the paper addresses a real gap in weather-aware multimodal 3D detection. Discrete weather categories give only a coarse description of environmental conditions and need not match the type, severity, or spatial pattern of the actual sensor degradation. Existing approaches also depend on camera-based weather recognition, which can itself become unreliable in adverse visual conditions. Deriving routing cues directly from LiDAR and 4D Radar observations is therefore a reasonable and practically relevant direction. It also allows DOME to perform comparably to camera-assisted methods without the camera modality.

Second, the PDR design is well motivated, and the paper examines it with analyses that go beyond overall detection accuracy. The global branch captures scene-level multimodal context, while the local branch models LiDAR–Radar disagreement, reflecting degradation that affects one modality more than the other. The component ablation supports this complementarity: global-only and local-only variants reach 59.7 and 58.8 AP3D, compared with 61.6 for the full PDR. In addition:

Sample-adaptive routing outperforms both weather-fixed routing (60.2) and image-based weather routing (59.4). In the controlled fog experiment, scene content and Radar observations are held fixed, and routing divergence increases with fog severity. In the held-out snow experiment, performance drops by only 2.6 and 2.5 points, compared with 15.8 and 14.9 for L4DR.

The lightweight DOME-L variant (about 78M parameters) and the InterFusion experiments further suggest that the approach is practical and not tied to a single backbone.

Critical weaknesses

My main concerns are that (i) the source of the improvement is unclear, and the contribution attributed to the proposed distortion-aware routing is small and confounded with MoE capacity and training strategy, and (ii) DOME does not outperform existing weather-aware methods on the main benchmark. Together, these issues lead to my current rating.

First, it is unclear what drives DOME's improvement. Most of the 8.1-point AP3D gain over L4DR comes from expanding to a generic MoE with about five times more parameters (53.5 → 59.7). However, the paper includes neither a parameter-matched dense L4DR nor a single conditional model that uses the same PDR fingerprint without multiple experts. As a result, it remains unclear whether the MoE formulation itself is necessary. The remaining 1.9-point gain attributed to distortion-aware routing is also difficult to interpret. Generic-MoE (Table 3) and Global-only PDR (Table 8) both reach exactly 59.7, which suggests that this gain largely corresponds to adding the local LiDAR–Radar discrepancy branch. It is also not stated whether Generic-MoE was trained with the same Progressive Expert Maturation strategy. Without it, PDR+DSE reaches only 58.3, below Generic-MoE. Finally, this comparison is reported from a single run. This is consequential because the paper's central claim, that distortion-aware routing drives robustness, is not cleanly supported by the current evidence.

Second, the comparison with prior weather-aware methods is less favorable than the text suggests. In Table 1, L4DR-DA3D achieves a higher overall AP3D than DOME (61.9 vs. 61.6), and AW-MoE achieves a higher overall APBEV (84.2 vs. 83.2). Under Rain, DOME is notably weaker (58.2 vs. 63.0 and 63.5). The augmentation-matched comparison in Appendix B.6 uses a non-standard protocol: it fine-tunes the detector with precomputed, fixed routing scores, so the 62.7 vs. 61.5 comparison is not fully matched. On OmniHD-Scenes, the improvements over L4DR are around 1 AP, and AW-MoE is not included. This is consequential because the practical value of DOME then rests mainly on achieving comparable performance without the camera modality, which is a weaker claim than the state-of-the-art performance the paper suggests.

Questions for authors
  1. Why is a Mixture-of-Experts formulation necessary for this problem?

The paper argues that predefined weather categories are inadequate because physical distortions vary continuously in type and severity. However, why should this continuous variation be represented through a finite set of experts rather than through a single condition-adaptive model?

Could the authors compare DOME with a parameter- and FLOP-matched universal model that receives exactly the same PDR distortion fingerprint but uses it to continuously condition a shared network, for example through feature modulation, dynamic residual adapters, conditional normalization, or another lightweight conditional transformation? Could they also report a dense L4DR scaled to a comparable parameter count? These baselines would separate the contribution of the distortion representation from the architectural benefit of introducing multiple expert branches, as well as from a simple increase in parameter count.

  1. How should the gain from PDR be interpreted relative to the Generic-MoE and Global-only variants?

The Generic-MoE (Table 3) and Global-only PDR (Table 8) both achieve 59.7 AP3D. Could the authors clarify exactly how the generic router differs from the Global-only PDR variant? If they are nearly equivalent, should the 1.9-point gain be understood primarily as the contribution of the local discrepancy branch?

In addition, was Generic-MoE trained with the same ALE, TAS, and ABS components as DOME? PDR+DSE without PEM reaches only 58.3, which is below Generic-MoE. Could the authors report Generic-MoE with and without PEM, along with mean and standard deviation over multiple seeds for both Generic-MoE and DOME, so that the 1.9-point difference can be assessed against run-to-run variance?

  1. How should DOME's performance be positioned relative to prior weather-aware methods?

In Table 1, L4DR-DA3D achieves a higher overall AP3D and AW-MoE† a higher overall APBEV than DOME, and DOME is notably weaker under Rain. Could the authors discuss these cases, in particular why DOME underperforms under Rain? Regarding Appendix B.6, could the authors also report an augmentation-matched comparison in which DOME is trained end-to-end with UDMA (with PDR computing routing scores online) rather than fine-tuned with precomputed routing scores?

  1. Can the authors provide more direct validation that PDR learns physical distortion rather than correlated scene properties?

The current continuous-fog experiment is useful because it holds scene content and Radar observations fixed while changing LiDAR degradation. However, it evaluates only one corruption family at three severity levels.

Could the authors perform controlled same-scene experiments across multiple physically distinct corruption types and severities and test whether the learned fingerprint or routing distribution tracks known degradation variables? For example, reporting how well the fingerprint predicts or correlates with simulated attenuation, point-drop rate, noise magnitude, or another physically defined degradation parameter would provide more direct support for the "physical distortion fingerprint" interpretation. It would also help to report the JSD between routing distributions of different real weather categories as a reference scale for the values in Table 9.

  1. How well does DOME generalize to genuinely unseen degradation levels or distortion types?

The simulated fog experiments use the same severity values for training augmentation and evaluation, although the scenes themselves are disjoint. Could the authors evaluate interpolation and extrapolation to severity values that are never observed during training?

More importantly, could the authors perform a leave-one-weather-out experiment in which an entire degradation type is excluded from training and evaluated without test-time adaptation? This would provide stronger evidence that DOME discovers reusable latent degradation structure rather than specializing to the set of distortion regimes already observed during training.

  1. How does PDR respond to non-weather sources of cross-modal disagreement?

The local PDR branch explicitly treats LiDAR–Radar discrepancy as a useful distortion signal, but similar discrepancies may arise from calibration errors, temporal misalignment, object motion, random point dropout, or sensor-specific noise.

Could the authors report routing or fingerprint changes under controlled non-weather perturbations, rather than only downstream detection accuracy? This would clarify whether PDR is specifically sensitive to weather-induced degradation or more generally reacts to arbitrary cross-modal inconsistency.

  1. Can the authors provide stronger evidence for expert specialization?

The routing distributions show that different weather conditions activate experts differently and that there is no one-to-one mapping between weather categories and individual experts. However, it remains unclear what functional specialization each expert actually learns.

Could the authors perform expert-level interventions, such as disabling, swapping, or forcing individual experts and measuring the resulting degradation across corruption types and severity levels? Such an analysis would provide stronger evidence that the experts correspond to distinct latent degradation functions rather than largely interchangeable high-capacity branches.

  1. How should the efficiency of DOME-L be interpreted when training cost is included?

DOME-L substantially reduces inference-time parameter count and FLOPs, but it uses a two-stage training procedure: a full first-stage detector training (35 epochs) followed by an additional adaptation stage (15 epochs). Could the authors report total GPU-hours, peak training memory, and active parameters per input for L4DR, DOME, and DOME-L?

This would help distinguish parameter-efficient deployment from overall computational efficiency and make the practical trade-off clearer.

Reviewer_Za2cOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DOME, a distortion-driven Mixture-of-Experts framework for robust LiDAR–4D Radar 3D detection under adverse weather. Instead of routing experts according to discrete weather labels, DOME learns a distortion fingerprint from global multimodal context and local LiDAR–Radar discrepancies, and uses it to activate distortion-specialized experts. The paper also introduces DOME-L, which replaces full experts with LoRA-based experts for better efficiency. Experiments on K-Radar and OmniHD-Scenes show strong improvements, including an 8.1-point AP3D gain over L4DR on K-Radar.

Critical strengths

The problem formulation is well motivated. The paper identifies a meaningful limitation of weather-category-based routing: samples with the same weather label may exhibit very different sensor degradation. The proposed distortion-driven routing provides a more flexible alternative without relying on camera-based weather recognition.

The experimental validation is strong. In addition to the main K-Radar results, the authors include a capacity-matched Generic-MoE comparison, routing-cue comparison, continuously varying fog, held-out snow severity, a second dataset, and another fusion backbone. DOME-L also substantially reduces the parameter overhead with only a small accuracy drop.

Critical weaknesses

1.The “physical distortion” interpretation is still somewhat indirect. PDR is learned from global features and LiDAR–Radar feature discrepancies, which may also encode scene content, object distribution, or modality-specific representation differences. The synthetic fog experiment provides useful evidence, but it does not fully establish that the learned fingerprint corresponds to physical distortion rather than a broader domain representation.

2.Generalization to truly unseen degradation types is not fully demonstrated. The held-out experiments mainly remove one snow severity while retaining another, and the continuous fog experiment varies the intensity of a known degradation process. It would be more convincing to hold out an entire weather/degradation type during training and test whether the learned routing generalizes to it.

Questions for authors

1.Can the authors provide further evidence that the learned fingerprint captures physical sensor degradation rather than scene or semantic differences?

2.What happens if an entire degradation type, e.g. fog or rain, is excluded from training rather than only holding out one severity level?

3.How sensitive is DOME to the assumption that 4D Radar is the relatively stable modality? Would the routing remain effective when Radar itself is significantly degraded?

Reviewer_JdHkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to address the degradation of sensor performance due to weather conditions in multimodal 3D detection. The authors argue that pre-existing methods that learn a single underlying distribution to capture varying weather conditions have limitations and lead to distribution shifts. To address this, they propose DOME, a mixture-of-experts framework with a router that combines global multimodal features and local feature differences via cosine similarity, then activates a few expert branches at deployment. They further propose a LoRA variant, DOME-L, to reduce parameter overhead.

Critical strengths

Overall, the paper is well-written and clearly presented. I would argue that the authors identify an interesting research question: different weather conditions induce heterogeneous degradation patterns in the sensor. The experimental design in this paper is quite intuitive and insightful, separating the routing design from the expert capacity, and covers a reasonably broad set of applications.

Critical weaknesses

My primary concern is with the central conclusions drawn from the experimental results. Specifically, I interpret the results as supporting degradation-sensitive routing rather than identifying latent physical-distortion domains. In Section 3.2, the router learns generic feature discrepancies and global context through supervision signals from object detection. The signal itself does not explicitly encode any physical-distortion target or constraint. I interpret Figure 5 and the results in Table 9 as strong evidence for my speculation. The routing responds to corruption, which can be attributed to supervised learning, but increasing divergence from a clean input neither establishes that experts specialize in a particular physical mechanism nor shows that their representation is independent of scene context. My point here is that the generic router is not subjected to the same intervention and attribution diagnostics, so those diagnostics do not show which physical sensitivity is distinctive to PDR. I think this matters because physical-domain decomposition is the main conceptual claim, while the implemented ingredients are established MoE, cross-modal discrepancy, and LoRA mechanisms.

A follow-up question will be the training budget and generalization capability. In my experience, learning a router/gate from data can benefit from scaling but cannot guarantee generalizability across different situations. I see similar traces in your report: Appendix A.1 explicitly uses the same three fog intensities during training augmentation and evaluation, while Table 5 therefore tests new scenes at trained severities, rather than extrapolating to new fog intensities. In contrast, the held-out snow experiment provides stronger evidence of cross-severity generalization, but it compares against the smaller L4DR rather than the capacity-matched generic MoE. I am really curious how scaling with larger training budgets and generalizability correlate here, given that the gate isn't technically learning to identify explicit physical distortion, as I pointed out above.

Questions for authors

Key questions related to my primary concerns are as mentioned above. I have a few additional minor questions about the results. For Tables 3–4, were the training schedules, initialization, and checkpoint selection identical, and are repeated-seed results for the generic router available? In Table 6, how many training/test scenes and objects remain in each held-out snow setting, and was model selection performed without the excluded severity?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.