- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes MPLoRA-MoE for RGB-D salient object detection. The main idea is that depth isn't always reliable, so the model should learn how much to trust it at each feature scale. RGB and depth share a PVT-v2-B5 encoder, and each modality gets its own LoRA residuals. At each of the four scales, a small router softly mixes four fusion experts that take different combinations of raw and cross-aligned RGB/depth features. The router gets no reliability labels and is trained only through the saliency and edge losses. The fused features then go into an ESCNet-style edge-semantic decoder. On eight benchmarks, the method is best on ReDWeb-S, SIP, COME-E, and COME-H, and roughly on par with prior methods on NJU2K, NLPR, DUT-RGBD, and STERE. Ablations cover edge guidance, the number and type of experts, the routing temperature, and the LoRA rank.
Critical strengths
- Good cross-dataset results. The method is best on the four unseen benchmarks (e.g., ReDWeb-S S_alpha .773 vs. .752 for the best baseline), which is where adaptive fusion should matter most. This is the main reason I don't lean toward a clear reject.
- Table 3 is a useful control. Four heterogeneous experts beat four or eight homogeneous ones, so the gain is not just from adding branches. This partly supports the idea that different fusion paths are needed.
Critical weaknesses
- The central claim, that the router learns when to trust depth, is never shown. There is no plot of routing weights, no per-scale statistics, and no link to depth quality. Fig. 1(c) even shows depth-oriented routing at shallow scales and RGB-oriented routing at deep scales, which is the opposite of what the text says (L88-92, L463-465). Since the title and the framing rest on this claim, it's the main reason for my score.
- The gains can't be attributed to routing. The obvious ablation, the same four experts with uniform weights (no router), is missing, and so is an image-level router. The backbone is PVT-v2-B5 with no params/FLOPs comparison and no same-backbone baselines, so part of the gain may come from the encoder. Apart from the router, most parts are borrowed (LoRA, coordinate attention, deformable conv, the ESCNet decoder), so without these controls the contribution feels thin for ICLR.
Questions for authors
- What do the routing weights look like per scale and per dataset, and do they shift when depth is corrupted or dropped at test time? A clear yes would directly support the main claim and could change my score.
- How does the model do with uniform expert weights, or with a single image-level router?
- Can the authors compare with a few baselines using the same backbone and report params/FLOPs?
2 comments
Feedback
Thanks to the authors for carefully adding degradation experiments, routing comparisons and parameter statistics. Routing weights will change in the expected direction with mode degradation, which answered my most important question, so I raised the score to 3. However, except for routing, most of them are existing modules, and the incremental improvements of the routing mechanism compared to existing adaptive fusion methods (such as LAFB) are also limited. I still have reservations about this.
We have addressed your concerns with new experimental results in the main paper and additional details in the Appendix.
We thank the reviewer for highlighting the need for direct routing evidence, controlled ablations, and complexity reporting. We have added experiments and revised the relevant discussion.
Q1–Q2: Reliability-responsive and scale-wise routing (Tables 3–4; Appendix Table 6)
1. Controlled degradation (Table 3). We select 547 RGB-D pairs from eight benchmarks using independently trained RGB-only and depth-only F3Net models, retaining images with high S-measure in both modalities. The same images and masks are evaluated under clean, RGB-degraded, depth-degraded, and jointly degraded conditions (Appendix A.1).
Table 3 reports E1–E4 routing weights at all four scales. Averaged across scales, the dominant expert shifts from E1 (both-oriented, 0.415) on clean inputs to E3 (depth-oriented, 0.427) with degraded RGB, E2 (RGB-oriented, 0.428) with degraded depth, and E4 (all-source, 0.374) when both degrade. Responses also differ by scale: RGB degradation increases E3 by 0.283 at S1 vs. 0.149 at S4; depth degradation increases E2 by 0.202 vs. 0.170.
These results directly link controlled modality corruption to expert-weight redistribution. We interpret routing weights as task-relative fusion preferences, not calibrated depth-quality estimates. Figure 1(c) and the text now clarify that scale-wise routing does not impose a fixed shallow-to-deep modality preference.
2. Matched routing ablations (Table 4). With the backbone, four experts, and decoder fixed, we compare static scale-wise routing (learned input-independent weights), image-level dynamic routing (one shared distribution), and our scale-wise dynamic routing.
Their S-measure/MAE results are 0.906/0.043, 0.909/0.041, and 0.915/0.035 on SIP, and 0.844/0.078, 0.848/0.076, and 0.855/0.069 on COME-H. These results demonstrate benefits from input-dependent, scale-specific expert weighting beyond merely having four branches. The static control uses learned weights rather than a literal uniform mixture.
3. Expert structure (Table 5(a)). Four heterogeneous E1–E4 experts improve SIP S-measure from 0.909 to 0.915 over four homogeneous E4 experts while using fewer expert parameters (3.09M vs. 10.77M).
4. Corruption robustness (Appendix Table 6). On the same degraded images, our S-measure is 0.943 (RGB), 0.959 (depth), and 0.803 (both), versus the best competing results of 0.925, 0.953, and 0.788. This complements the routing statistics with predictive performance under matched corruptions. We test controlled degradation rather than complete depth removal or a severity sweep.
Q3: Backbone and computational complexity (Table 1; Appendix Table 7)
Table 1 now reports parameter counts; Appendix Table 7 additionally provides GFLOPs and input resolutions. MPLoRA-MoE has 105.17M parameters and 341.72 GFLOPs at 416×416, compared with 223.10M parameters for CPNet, 453.03M for LAFB, 272.10M for CATNet, and 368.42M for DPPNet.
All four experts execute under dense soft routing, so we do not claim computational savings from sparse expert activation. Different methods also use different backbones and resolutions; these comparisons are not backbone-controlled.
Although we have not retrained external baselines with PVT-v2-B5, Table 4 fixes our backbone and architecture across routing variants, isolating the routing strategy rather than encoder capacity.
We hope the new controlled routing statistics, matched ablations, and complexity results address the reviewer's core concerns.