acceptodds
Under review as a conference paper at ICLR 2027

Discarding Sensor Depth: Reliability-Aware Geometry Learning for Salient Object Detection

Abstract

RGB-D salient object detection (SOD) complements RGB appearance with scene geometry, but practical depth measurements are often noisy, incomplete, or structurally misleading, making tightly coupled RGB-depth models vulnerable to negative transfer. We propose \method, a depth-free RGB-D SOD framework that learns task-oriented geometry from RGB instead of consuming sensor depth. During training, a frozen foundation depth estimation model transfers dense relative structure, multi-scale feature organization, and boundary cues to an RGB-driven geometry branch, addressing how useful geometry is acquired without dataset depth. Cross-modal enhancement then exchanges appearance and geometry across scales, while reliability-aware fusion suppresses geometric responses that disagree with RGB evidence or are irrelevant to saliency, addressing how the learned geometry should be used. The external teacher is removed after training and sensor depth is never required by the final detector. Across nine RGB-D benchmarks, is best or tied-best in 26 of 36 metric–dataset comparisons, with the clearest gains on cross-dataset benchmarks. Results on four RGB SOD benchmarks further confirm that the learned geometry transfers beyond RGB-D sensor domains. Overall, task-adapted geometry retains the structural benefit of depth while avoiding direct dependence on unreliable measurements.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Salient Object Detection, RGB-D Salient Object Detection, Geometry Distillation, Depth-Free Learning, Reliability-Aware Fusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aVc9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Depth-free RGB-D. Uses a frozen depth estimation foundation model for transferring depth relative structure, multi-scale features and boundary information. Uses this for a geometry branch, then uses cross-modal enhancement. Finally, it is a distillation net, so distils this. Shows good results. Claims that RGB-D depth is not always high quality, and with repeated fusion can cause negative transfer. Main claims (non-results):

  1. Reformulate RGB-D SOD on RGB-Driven geometry learning avoiding depth-induced negative transfer at source rather than calibrating sensor depth after it enters detector.
  2. Multi-level distillation, cross modal enhancement and reliability aware fusion
  3. In related work under: reliable or unavailable depth - GeoDistill internalizes geometry in an RGB-Driven branch so neither sensor depth or external depth estimator is required at inference.
Critical strengths
  1. Results show good gains.
  2. The architecture is novel.
Critical weaknesses

Clarity on contribution and novelty of contribution. This paper makes many claims towards novelty, clouding the contribution. This makes it hard to pick apart what the actual contributions are. As such this review examines the claims specifically from the summary above.

3 &2) reliability aware fusion: This idea is well-established. For example, fully-published before this conference due date: Channel-Wise Contribution Assessment for RGB-D Salient Object Detection, Cheng, et al, ACM Transactions on Multimedia Computing, Communications and Applications, Volume 22, Issue 8, 2026. Cheng et al directly uses reliability aware fusion for RGB-D. More generally, This type of reliability/uncertainty/error driven training idea in SOD/COD has a long history, for example, Modeling aleatoric uncertainty for camouflaged object detection, Liu, et al, WACV 2022. 2) Distillation itself in SOD is not novel and is well-known: eg., Self-distillation salient object detection via generalized diversity loss, Yunfei Zheng, Pattern Recognition, 2025.

1 & 4) The core claim of the paper is around using RGB-D at training time and RGB only at test time with some type of distillation. This has been commonly done with train time only branches. This is a long-established topic, early papers that have depth supervision at training time to improve the performance of RGB only SOD at test time include: RGB-D Salient Object Detection with Ubiquitous Target Awareness, Y Zhao 2021, ARXIV; and, Is depth really necessary for salient object detection? Zhao et al, ACM Multimedia 2020.

The difference that the authors seem to be drawing towards is that the network does not supervise using the depth directly. The implication is that the depth supervision is not direct. Here in the literature review the authors contrast with SATNet. Where the estimated depth map enters as an explicit modality. The difference claimed here in Fig 1 is to distil and use geometry teaching rather than use Depth Ground truth and learn depth estimation. Yet depth is shown as a direct training input from Depth Anything V2 in the top of Fig 2, and the network directly uses Depth Anything V2 features. This is slightly indirect depth supervision – the difference is not very strong from above. Yes, the network also uses distillation and edges (Taken directly from Ye 2025), but this does not appear to be the only place where depth features are available directly to the network. So although the Geometry branch trains RGB based features, they are supervised directly from a depth network at training time (not conceptually different to Zhao et al 2020 and Zhao et al 2021. So rather than a conceptual difference this difference seems mostly an internal architectural difference as the inputs are the same, and overall the network has direct access to depth feature supervision at training time.

Questions for authors

The key question is whether this is a misunderstanding of the paper, that there is a strong conceptual difference from the background literature in terms of the manner in which depth supervision (estimated or otherwise) is being used at training time to distill an RGB only network for SOD at test time. Or whether it is an architectural difference. I cannot see evidence that there is a strong conceptual novelty from prior work in SOD.

3 comments
AuthorsOpenReview ↗

Clarifying the Core Novelty and Providing Evidence for Task-Adaptive Geometry Distillation

We thank the reviewer for this critique. We agree that depth-assisted training with RGB-only inference (Zhao et al., 2020, 2021), distillation, reliability-guided fusion, and edge-aware decoding are not individually new. Our presentation blurred this distinction. We do not claim to invent these ingredients or RGB-only inference. We instead focus on acquiring and using geometry for SOD without dataset sensor-depth supervision.

1. What differs from prior depth-assisted SOD?

Zhao et al. (2020, 2021) use measured depth from RGB-D training sets as supervision and remove it at inference. Our SOD training uses RGB-mask pairs, not their measured depth. A frozen Depth Anything V2 (DA-V2) model processes RGB and provides relative-depth outputs, intermediate encoder features, and boundaries as training targets. This is depth-related supervision, not learning without prior depth knowledge. The difference is the source and form of supervision: pretrained geometric priors rather than measured target-dataset depth labels.

SATNet substitutes Depth Anything pseudo-depth but still feeds an estimated depth map to its detector. GeoDistill internalizes teacher knowledge in an RGB-driven geometry branch; neither sensor depth nor a running depth estimator is needed at inference. Retraining on RGB-only DUTS-TR further demonstrates this flexibility.

In Fig. 2, the frozen teacher receives RGB; the depth shown is its prediction, not observed depth fed to the student. Fig. 1 contrasts measured-depth and foundation-teacher supervision, not the absence of all depth-derived knowledge.

2. Why is this more than swapping depth labels?

Our geometry branch learns from complementary output, feature, and boundary targets while being optimized for SOD. Feature interaction and a learned gate regulate where geometry influences saliency. Neither distillation nor gating is claimed as a new primitive.

The cited 2026 channel-wise contribution work regulates observed RGB-D modalities, whereas our fusion regulates RGB-derived geometry. We also acknowledge that AETP/ESC decoding primitives originate from Ye et al. (2025).

Our contribution is the task-adaptive acquisition and utilization of foundation geometry without target-dataset depth, not a new theory of distillation, uncertainty, or fusion.

3. What experimental evidence supports this distinction?

Revised Table 4 isolates five design choices:

(a) Supervision source: On SIP, DA-V2 output supervision improves SmS_m from .886 (unsupervised geometry branch) to .893, whereas measured GT-depth supervision reaches .887. This supports foundation geometry over directly reproducing measured depth.

(b) Knowledge representation: With DA-V2 fixed, feature-only distillation improves SIP SmS_m/MAE from .893/.042 (output-only) to .907/.036; combining both reaches .912/.033. Thus, our transfer is more effective than simply reproducing pseudo-depth.

(c) Teacher knowledge: Under matched feature-only distillation, DA-V2 encoder features outperform plain DINOv2 on SIP (SmS_m: .907 vs. .905; MAE: .036 vs. .039), also improving DUT-RGBD and COME-H. This supports depth-specialized representations beyond generic visual features, although pretraining data differences may contribute.

(d) Task adaptation: Direct DA-V2 feature fusion is weaker than geometry-student distillation (SIP SmS_m: .898 vs. .912), demonstrating the benefit of adapting teacher representations to SOD.

(e) Geometry utilization: Replacing reliability-aware fusion with Concat increases SIP MAE from .033 to .038, validating our fusion design without claiming conceptual novelty for reliability gating.

The new depth-cue analysis further clarifies our scope. On the 1,000 samples with the least useful depth cues, GeoDistill achieves SmS_m=.752 versus .719 for the strongest RGB-D competitor. Conversely, RGB-D methods slightly outperform ours when depth is informative. This supports avoiding dependence on unreliable or uninformative depth, not assuming all sensor depth is harmful.

In summary, our novelty is not RGB-only inference, distillation, reliability gating, or decoder primitives individually. It lies in replacing target-dataset measured-depth supervision with foundation-guided, SOD-adapted geometry and selectively utilizing it without external depth at inference. The new supervision, teacher, transfer, and fusion controls substantiate this method-level contribution. We will revise our contribution statements and figure labels accordingly.

Reviewer_aVc9OpenReview ↗

Summary?

So, not RGB-only. Using Depth Anything V2 within the learning loop as SgFx points out. Not using explicit depth as explicit supervision, or using foundation model generated depth directly as supervision, but using Depth Anything V2 within the learning so that foundation model generated depth is available within the overall learning pipeline? The difference you point to is using "geometry" - features from the foundation model, rather than its output, which is depth? - "replacing target-dataset measured-depth supervision with foundation-guided, SOD-adapted geometry".

Also, as SgFx points out, always ignoring sensor depth as the premise, rather than considering where to replace it? -

Thank you for following up. We agree our wording blurred depth-guided training and RGB-only inference.

1. What enters the learning pipeline? DA-V2 is a frozen, training-only teacher. During SOD training, it takes RGB—not measured depth—as input and provides four encoder features, predicted relative depth, and derived boundaries. The student uses both feature distillation and auxiliary depth/boundary supervision. The depth shown in Fig. 2 is the teacher's prediction, not sensor input. Neither measured depth nor DA-V2 is needed at inference. Our setting is no target-dataset sensor-depth supervision and RGB-only inference, not training without depth-derived knowledge.

2. Are “geometry features” simply depth features? We claim neither a new feature type nor that DA-V2 encoder features contain geometry exclusively. “Geometry” denotes depth-specialized structural knowledge adapted to SOD. The final model combines feature distillation with depth-output supervision; it does not replace the latter entirely. Table 4(b) shows that, with DA-V2 fixed, feature-only exceeds output-only supervision on SIP (SmS_m: .907 vs. .893; MAE: .036 vs. .042). Under matched feature-only settings, DA-V2 also exceeds plain DINOv2 (SIP MAE: .036 vs. .039), supporting the usefulness of depth-specialized representations. Crucially, direct DA-V2 feature fusion underperforms the task-adapted geometry student (Table 4(d), SIP SmS_m: .898 vs. .912). The contribution is therefore not merely accessing DA-V2 features, but transferring and adapting them for saliency. We do not claim distillation itself is new.

3. Why always omit measured depth? This is a deployment constraint, not a claim of universal optimality. Selective RGB-D fusion is valid when depth is available, but requires that input and judgments about its usefulness. We study the complementary setting without these requirements. Our new analysis reveals the trade-off: on Worst-1000, GeoDistill achieves SmS_m=.752 versus .719 for the strongest RGB-D baseline; on high-usefulness subsets, RGB-D methods slightly outperform ours. We do not claim ignoring reliable depth is always better.

Our contribution is a tested method for learning SOD-adapted representations from depth-foundation guidance without measured depth in target SOD datasets, not a new RGB-only inference paradigm or definition of geometry. We will clarify this and Fig. 2. We are confident that our work makes a meaningful contribution to the field and will inspire further research into sensor-depth-independent, task-adaptive geometry learning for SOD.

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

GeoDistill learns geometry for salient object detection without using dataset sensor depth during training or inference. A frozen Depth Anything V2 teacher supervises an RGB-driven geometry branch through dense predictions, intermediate features, and boundaries. The learned geometry interacts with appearance features and enters the saliency decoder through gated fusion. The teacher is removed at inference. Experiments cover nine RGB-D benchmarks and four RGB benchmarks, with ablations on supervision sources, feature transfer, and model capacity.

Critical strengths

1.The results are convincing across several datasets. GeoDistill is best or tied-best in 23 of 24 comparisons on the six RGB-D benchmarks outside its training mixture. The ReDWeb-S MAE improvement, from 0.112 to 0.097, is particularly clear.

2.Table 4 goes beyond simply adding modules. It compares an RGB-only model, an unsupervised geometry branch, sensor-depth supervision, and teacher supervision. Feature distillation consistently beats output-only supervision.

3.The final model needs neither sensor depth nor the external teacher. The separate RGB SOD experiments also show that the approach is useful beyond the RGB-D setting.

Critical weaknesses
  1. The reliability gate is not separately evaluated, despite being a central part of the method. Its benefit over ordinary learned fusion is unclear.

2.The teacher uses DINOv2 features. Without a plain DINOv2 control, it is difficult to tell how much of the gain comes specifically from depth learning.

3.The explanation of the cross-dataset gains is stronger than the evidence. Better generalization is demonstrated, but sensor-depth noise is not isolated as its cause.

4.Baseline evaluation details are sparse. PICRNet’s consistently low scores need checking, although the method also outperforms much stronger baselines.

Questions for authors
  1. How much performance is lost if the reliability gate is replaced with simpler fusion?
  2. Have you compared the depth-trained teacher with plain DINOv2 under the same feature-distillation setup?
  3. Were the baseline results obtained from released predictions or rerun models? Please explain the checks on PICRNet.
  4. In Table 4(c), what outputs and losses are used for YOLO26-X, and how do they correspond to the DA-V2 setup?
  5. Which validation data were used to choose the projection width and supervision configuration?
1 comment
AuthorsOpenReview ↗

Additional Ablations, Baseline Corrections, and Experimental Protocol Clarifications

We thank the reviewer for the constructive feedback. We added targeted teacher and fusion ablations, a depth-stratified evaluation, and a corrected PICRNet assessment, and clarified our experimental protocol.

1. Reliability-aware fusion (Q1). Table 4(e) compares our fusion with Add (element-wise summation followed by a 3×3 convolution) and Concat (channel concatenation followed by 1×1 and 3×3 convolutions). All variants retain the same RGB encoder, geometry branch, feature enhancement, distillation objectives, and training schedule; only the fusion operator changes. Our method leads on all 12 metrics across DUT-RGBD, SIP, and COME-H. Against Concat, MAE drops from .024 to .020 (16.7%), .038 to .033 (13.2%), and .067 to .065 (3.0%), respectively. Thus, learned reliability-aware regulation improves over ordinary learned fusion. We do not claim the gate estimates calibrated sensor-depth reliability.

2. Plain DINOv2 control (Q2). We added a frozen, original DINOv2-S encoder teacher, using the same four encoder layers and spatial feature-alignment loss as the DA-V2 teacher. Both have the ViT-S/14 encoder architecture, so the comparison controls for backbone capacity and feature-extraction stages. Against DA-V2 feature-only distillation (Table 4(b)), plain DINOv2 obtains SmS_m=.944/.905/.848 versus .949/.907/.857, and MAE=.022/.039/.073 versus .020/.036/.066 on DUT-RGBD/SIP/COME-H. The full DA-V2 configuration further achieves SmS_m=.948/.912/.856 and MAE=.020/.033/.065. The matched feature-only comparison supports the value of depth-specialized representations beyond generic DINOv2 features; we recognize that DA-V2's additional pretraining data may also contribute to the difference.

3. Cross-dataset gains and depth-cue usefulness. We agree that cross-dataset performance alone cannot establish sensor noise as its cause. We trained depth-only F3Net on the standard 2,985 RGB-D training pairs (using depth as input) and ranked images pooled from all nine test benchmarks by depth-only SmS_m. The highest/lowest 500 and 1,000 images form four diagnostic subsets. On Worst-1000, ours achieves SmS_m=.752 and MAE=.119 versus .719/.133 for the strongest RGB-D comparator, and leads all four metrics on both worst subsets. Conversely, on the best subsets, RGB-D models modestly outperform ours, confirming that informative measured depth remains beneficial. This is not an all-or-nothing claim: our method removes dependence on sensor depth rather than proving it should always be discarded. Depth-only SmS_m measures depth-cue usefulness for SOD, not physical sensor accuracy; RGB difficulty and dataset composition may also influence these trends. We have moderated the causal interpretation accordingly.

4. Baselines and PICRNet (Q3). For each baseline, we obtained code from the official author-provided link. When released checkpoints were available, we ran local inference and evaluation with those weights; otherwise, we retrained using the official implementation and paper-specified settings, without modifying the architecture. PICRNet's original anomalous scores resulted from an incompatible newer Detectron2 version that caused incorrect computation. After installing its required compatible version and rerunning inference, we corrected all nine benchmark results (e.g., NJU2K SmS_m: .386→.899; SIP: .312→.872; COME-H: .361→.808). We checked the remaining baselines for analogous issues and found none. Comparisons with the strongest competitors, and our central conclusions, remain unchanged. We appreciate the reviewer for drawing attention to this error.

5. YOLO26x-Depth and model selection (Q4–Q5). Table 4(c) uses YOLO26x-Depth, a monocular metric-depth estimator, not a YOLO26 object detector. Its intermediate feature supervision follows the same feature-alignment procedure and loss configuration as the plain DINOv2 control; the full DA-V2 configuration additionally uses normalized relative-depth output and boundary supervision. We have specified these differences to avoid interpreting Table 4(c) as an exclusively feature-only teacher comparison. For projection width and supervision configuration, we empirically evaluated candidate choices over all nine benchmark test sets and selected the best overall-performing setting; the three datasets displayed in the ablation tables are a subset of that assessment. No independent held-out validation split was used for this choice. We now state this explicitly and acknowledge the potential selection bias from using test results.

Together, the new controlled comparisons strengthen the evidence for the fusion design and depth-specialized knowledge transfer; the corrected baseline results and qualified generalization claims make the evaluation more transparent.

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

The authors argue that the conventional paradigm of using sensor-acquired depth maps in RGB-D salient object detection (SOD) may introduce noise and misleading cues that degrade model performance. To address this issue, they propose GeoDistill, a method that eliminates the need for sensor depth during inference and instead leverages a powerful pretrained depth estimation model, Depth Anything V2, to provide supervision during training. This enables the model to learn task-relevant geometric information from RGB inputs, which is then integrated with RGB features through cross-modal enhancement for salient object segmentation. Evaluations on nine RGB-D benchmarks show that GeoDistill achieves state-of-the-art or tied-best performance.

Critical strengths
  1. The authors revisit the role of the depth modality in RGB-D salient object detection and challenge the conventional way of utilizing depth information. Instead, they leverage pseudo-depth generated by a pretrained depth estimation model to inject geometric cues into the detection task.

  2. Through reliability-aware perceptual fusion, GeoDistill effectively integrates geometric and RGB information, leading to improved segmentation performance.

  3. The authors introduce a geometric knowledge distillation approach that successfully transfers knowledge from a pretrained depth estimation model to an RGB-based network.

Critical weaknesses
  1. Questionable motivation. A substantial body of prior work on RGB-D salient object detection has demonstrated the importance of depth information, particularly in challenging and complex scenarios. Depth cues are also important for distinguishing relative saliency between objects and their surroundings. More precisely, determining when depth is beneficial and when it becomes detrimental is a meaningful practical problem. However, the authors appear to generalize the potential negative effects of depth to all scenarios and consequently eliminate sensor depth entirely. This framing overstates the problem and does not adequately justify the proposed all-or-nothing solution.

  2. Insufficient experimental analysis. The authors do not provide sufficient empirical evidence to demonstrate that real-world depth measurements frequently suffer from noise, incompleteness, or misleading structural information, nor do they systematically analyze how these issues affect detection accuracy and salient object perception.

  3. Unclear source of performance gains. It remains unclear whether the observed improvements primarily stem from the proposed geometric knowledge distillation mechanism itself or from the pretrained knowledge provided by Depth Anything V2.

Questions for authors
  1. The authors have not sufficiently demonstrated that discarding sensor depth is a better solution than improving depth fusion. Previous studies have already established that depth information can be beneficial in certain challenging scenarios. The authors should therefore explain why depth should be discarded entirely rather than selectively or adaptively utilized.

  2. Does discarding sensor depth remain advantageous when the depth measurements are reliable? Does the proposed method outperform adaptive fusion strategies designed to handle unreliable depth inputs? The authors should provide additional robustness analyses and ablation studies to address these questions.

  3. Since the proposed method introduces Depth Anything V2, a powerful pretrained depth estimation model, as an additional source of supervision, the authors should disentangle the contributions of the proposed geometric distillation mechanism and the pretrained knowledge supplied by Depth Anything V2. Additional controlled experiments are needed to establish whether the improvements can be attributed specifically to geometric distillation.

3 comments
AuthorsOpenReview ↗

Clarifying the Motivation and Validating Robustness Across Different Depth-Cue Quality Levels

We thank the reviewer for highlighting the value of reliable depth and for requesting a more precise motivation and controlled evidence. We agree that depth can substantially improve SOD, especially when RGB appearance is ambiguous. Our claim is not that all sensor depth is poor, nor that discarding reliable depth always yields higher accuracy. The central issue is that the usefulness of geometry and the reliability of a particular depth measurement are different: missing regions, foreground–background bleeding, and misleading discontinuities can become plausible encoded features, making their influence difficult to eliminate solely through feature-level fusion. We therefore study a complementary setting: can an RGB-driven detector acquire useful geometric priors without requiring dataset sensor depth during SOD training or depth input at inference? This avoids dependence on depth availability and quality; it is not a claim that adaptive RGB-D fusion is unnecessary when reliable depth is available. We have clarified this scope in the revised manuscript.

Depth usefulness and robustness (Questions 1–2).

We added a stratified evaluation specifically to test where our approach helps and where it does not. We trained a depth-only F3Net on the standard 2,985 depth–mask training pairs, then ranked images pooled from all nine RGB-D benchmarks by the depth-only S-measure. The bottom/top 500 and 1,000 images form four subsets (Table 5); these scores measure depth-cue usefulness for SOD rather than physical sensor accuracy.

On the Worst-1000 subset, GeoDistill achieves Sm=.752S_m=.752 and MAE=.119, versus Sm=.719S_m=.719 and MAE=.133 for the best conventional RGB-D baseline. It leads CPNet, CATNet, and SATNet on all four metrics for both low-usefulness subsets. Conversely, on the Best-1000/500 subsets, RGB-D baselines slightly outperform GeoDistill, consistent with the reviewer's observation that reliable and informative depth can complement RGB.

Thus, the new results directly reject an "all-or-nothing" interpretation: our advantage is strongest when observed depth is unhelpful, not when it is consistently informative. Table 1 additionally compares with established fusion methods, including CAVER and LAFB; GeoDistill is best or tied-best in 23/24 comparisons on six unseen-source benchmarks. Our fusion ablation (Table 4e) also shows that replacing reliability-aware fusion with Concat increases MAE from .033 to .038 on SIP and from .020 to .024 on DUT-RGBD.

We also acknowledge the limit of this analysis. Poor depth-only saliency may reflect sensor defects, weak depth/saliency correspondence, difficult RGB scenes, or dataset differences; notably, 859 of the Worst-1000 images come from COME-H, COME-E, and ReDWeb-S. Hence we no longer interpret cross-dataset gains or these subsets as causal proof that sensor noise alone explains the improvement. Rather, they show a consistent performance advantage under low depth-cue usefulness, while exposing the setting where direct RGB-D fusion remains beneficial.

Distillation versus pretrained teacher knowledge (Question 3).

The revision adds complementary controls in Table 4.

First, compared with a geometry branch without teacher supervision, DA-V2 output supervision improves all 12 reported metrics; supervision from measured GT depth is weaker in 11/12 comparisons (Table 4a).

Second, with the same student design, feature-only distillation outperforms output-only supervision on all 12 metrics: on SIP, SmS_m increases from .893 to .907 and MAE decreases from .042 to .036 (Table 4b).

Third, a plain DINOv2 encoder control addresses generic visual pretraining: using the same feature-alignment mechanism, the DA-V2 teacher improves SmS_m from .905 to .912 and MAE from .039 to .033 on SIP, with improvements on DUT-RGBD and COME-H as well (Table 4c).

Finally, directly fusing DA-V2 teacher features is worse than training the RGB-driven geometry student (SIP SmS_m: .898 vs. .912; Table 4d).

These comparisons separate, to the extent feasible, the value of depth-specialized teacher representations from the effectiveness of our task-adaptive distillation and student learning. We do not claim that the DA-V2 prior has no contribution; rather, our method benefits from that prior and demonstrates that how it is transferred and used for SOD matters.

In short, our goal is not to replace informative sensor depth in every scenario. It is to learn transferable, saliency-adapted geometric structure without depending on measured depth, and the new stratification and ablations clarify both its benefits and its limits.

I have read the author responses, but unfortunately my opinions and concerns persist: 1) The motivation of discarding sensor depth is not always reasonable. The authors have clearly confirmed this in the rebuttal. This poses a concern that a better way should determine when to discard sensor depth and when not, but the method has not realized this. 2) Continued after 1), if the method always discard sensor-depth, the method is more like an rgb-based method (only rgb input is need for all pipeline including the Depth Anything V2 model), and so placing it in an rgb-d track is not that favorable. 3) As reviewer aVc9 raised, the distillation idea is not that new and a few previous methods have already explored depth supervision at training-time-only. The major difference seems this work employs Depth Anything V2 model to generate the depth while previous works use the sensor depth, but this again leads to concerns 1) and 2).

We thank the reviewer for the follow-up. Reliable sensor depth can help, but this does not invalidate methods that avoid depending on it. We clarify three points.

1. Why not selectively discard sensor depth?

Selective fusion is a valuable direction, but it requires reliable decisions at test time, without ground-truth depth quality or an oracle modality choice. This difficulty has already been examined in RGB-D SOD. Fan et al., Rethinking RGB-D Salient Object Detection (TNNLS 2021), introduced D³Net's Depth Depurator Unit (DDU), which chooses between RGB and RGB-D predictions to filter low-quality depth. Their Table V shows that the learned selector does not consistently improve over the RGB-D branch (e.g., SIP S-measure .860 vs. .862; LFSD .825 vs. .836) and remains about 1.6 percentage points below an oracle selector on average. This does not discredit selective fusion; it shows that accurate selection is nontrivial.

Moreover, image-level quality and pixel-level complementarity are not equivalent. A globally poor depth map may still contain useful local boundaries, while an accurate depth map may contain structures irrelevant to saliency. Ideal fusion must determine both where depth is reliable and where it adds information beyond RGB. Discarding an entire map can lose local cues; retaining it can propagate misleading responses. GeoDistill instead distills geometry without needing sensor depth or reliable quality judgments at deployment. It does not claim to dominate an oracle adaptive-fusion system.

Our new stratified evaluation supports this scope. Using depth-only F3Net to rank samples pooled from all nine RGB-D benchmarks, GeoDistill achieves S-measure .752/MAE .119 on Worst-1000, versus .719/.133 for the strongest RGB-D baseline, and leads across all four metrics on both low-usefulness subsets. On high-usefulness subsets, RGB-D methods perform slightly better. These strata reflect depth-cue usefulness, not measured sensor noise; they show an advantage under weak depth cues, not universal superiority.

2. Is this merely an RGB method outside RGB-D SOD?

We agree that the deployed detector is RGB-only in its input requirements. This is intentional. But it is depth-knowledge-guided: DA-V2 was trained to predict depth, and its frozen representations and depth-derived targets supervise our geometry branch during SOD training. Thus, sensor depth is unnecessary in the target SOD datasets, whereas geometric knowledge still originates from depth-supervised foundation learning. This makes the work relevant to RGB-D SOD's central question of how to exploit geometry when measured depth is absent or unreliable; we compare under established RGB-D splits precisely to evaluate that alternative. We also evaluate on four standard RGB SOD benchmarks to show transfer beyond RGB-D datasets. We will explicitly describe it as depth-guided training with RGB-only inference.

3. What differs from prior training-time depth supervision?

We agree that distillation and RGB-only inference are established, and that simply replacing sensor depth with DA-V2 would be insufficient novelty. Our focus is on what representation is transferred and how it is adapted to saliency. Rather than relying primarily on regression to a final measured or pseudo-depth map, we align four intermediate DA-V2 encoder representations with an RGB-driven geometry student, train that student jointly for SOD, and selectively integrate its geometry. Output and boundary supervision remain complementary.

The ablations isolate why these choices matter: with DA-V2 fixed, feature-only distillation exceeds output-only supervision on SIP (S-measure .907 vs. .893; MAE .036 vs. .042); direct DA-V2 feature fusion is weaker than the task-adapted student (.898 vs. .912 S-measure); and a matched plain DINOv2 feature-only control underperforms DA-V2 (.905 vs. .907 S-measure; .039 vs. .036 MAE). The fusion ablation separately favors reliability-aware fusion over Concat (SIP MAE .033 vs. .038). We do not claim these primitives are individually new. Together, they show that depth-pretrained information must be selected, transferred, and adapted for SOD—not simply supplied as a depth map or raw teacher feature.

In short, selective sensor-depth fusion and sensor-depth-independent geometry learning address different constraints and can coexist. Our contribution is a validated method for the latter, not the assertion that reliable depth should always be discarded. We will make this positioning explicit in the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.