acceptodds
Withdrawn from ICLR 2027

PhysCorr: Reliability-Controlled Correction for Visible-Infrared Camouflaged Object Segmentation in Unseen Environments

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Visible–infrared camouflaged object segmentation (VICOS) is vulnerable in unseen environments, where the two modalities often suffer asymmetric degradation. Existing fusion methods rely on statistical complementarity and may propagate corrupted modality cues, while visual structures can be overwritten by unconstrained physical priors. To address these limitations, we propose PhysCorr, a reliability-controlled physical correction framework that selectively refines rather than replaces visual evidence. Specifically, imaging conditions are factorized into structured degradation states and target-invariant properties. Shared hierarchical features establish a stable visual reference, while residual correction is generated under physical-state modulation. A reliability-conditioned intervention mechanism integrates cross-modal agreement, branch reliability, spatial uncertainty, and boundary evidence to determine when, where, in which direction, and to what extent correction is applied. Counterfactual invariant learning aligns target representations and invariant reasoning states between observed conditions and clean counterfactuals, suppressing environment-specific shortcuts. Consequently, stable predictions are preserved under conventional conditions, while physical correction is adaptively strengthened in adverse environments. Experiments under conventional and cross-environment settings show that PhysCorr improves source-domain segmentation accuracy and robustness to unseen asymmetric degradation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visible-Infrared Camouflaged Object Segmentation, Multimodal Learning, Physics-guided Learning, Reliability-aware Fusion, Domain Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qVshOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PhysCorr, a reliability-controlled physical correction framework for VICOS in unseen environments. PhysCorr models degradation as structured states, selectively applies physical correction based on cross-modal reliability, and promotes environment-invariant representations through counterfactual learning. Experiments demonstrate the effectiveness of PhysCorr.

Critical strengths
  1. The proposed paradigm that physical knowledge should act as adaptive corrective evidence rather than replace reliable visual observations is well-motivated and could promote further research interest in complex scene perception.
  2. This paper is generally well-written and easy to follow.
Critical weaknesses
  1. PhysCorr achieves only marginal improvements over the state-of-the-art method in Table 1. Given that PhysCorr involves multiple hyperparameters, including those in Eqs. (9) and (15), even small changes in their values may substantially affect its performance. This raises concerns about the robustness of the reported gains and the effectiveness of the proposed corrective paradigm.
  2. Although Qwen remains frozen throughout the process, PhysCorr’s use of the 7B-parameter model increases the overall parameter count and inference overhead, potentially limiting the method’s practical applicability.
  3. How would the experimental results change if the Qwen outputs used in Eqs. (2) and (3) were replaced with dimensionally consistent random values?
Questions for authors

Please refer to Weaknesses.

Reviewer_fiEmOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PhysCorr, a reliability-controlled physical correction framework for visible-infrared camouflaged object segmentation in unseen environments, which aims to address the problem that traditional fusion methods tend to propagate erroneous information when different modalities are asymmetrically degraded.

The method distinguishes between environmental degradation factors and target-invariant properties through structured physical state modeling, and generates residual corrections through physical state modulation based on multimodal visual features. By combining cross-modal agreement, branch reliability, spatial uncertainty, and boundary information, the location and strength of physical correction are adaptively controlled to prevent erroneous physical priors from damaging reliable visual structures. The introduction of counterfactual invariant learning enhances the consistency of target representations under different degradation conditions.

Critical strengths

Through the residual correction mechanism, the paper uses physical priori as auxiliary information instead of directly replacing visual prediction, which helps to reduce the interference of unreliable physical information on the segmentation results.

The paper focuses on the asymmetric degradation of visible and infrared modal in the unseen environment, which is of practical significance.

Critical weaknesses

The method uses a frozen Qwen2.5-VL to generate physical states, but their physical accuracy and causal relationships have not been sufficiently verified, especially in unseen environments.

Although the frozen VLM generates a physical causal chain, the reasoning itself does not directly participate in segmentation. It remains unclear whether the generated states truly reflect physical properties or simply serve as additional semantic information.

The method uses physically guided field expansion during training, while the evaluation mainly focuses on predefined fog, forest, and snow environments. It is unclear how much of the improvement in cross-environment generalization comes from the proposed physical reasoning rather than data augmentation.

Some concepts in the method section could be better defined and explained. The motivation behind certain design choices also needs to be clarified, with further analysis to support their effectiveness.

Questions for authors

The physical states seem to be limited to the degradation types considered in the paper. I wonder whether the method would still work when different degradations occur together or when it encounters degradation types not seen during training.

I am not sure how a single low-dimensional physical state can describe the degradation across an entire image. If some regions are more severely degraded than others, how does the model determine the appropriate correction for each region?

Reviewer_P3GEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhysCorr for visible-infrared camouflaged object segmentation under unseen environmental conditions. It uses structured physical states generated by a vision-language model to guide reliability-controlled residual correction. Counterfactual invariant learning is further introduced to improve robustness to environmental shifts. Experiments on VIAC and EVIAC demonstrate improvements over existing methods in both conventional and cross-environment settings.

Critical strengths

S1: The proposed approach provides a novel and technically reasonable way to leverage environmental states inferred by a vision-language model to guide residual corrections, reducing the risk of errors caused by unreliable physical priors.

S2: Experiments on EVIAC show that PhysCorr improves segmentation performance under unseen environmental conditions, demonstrating its potential for cross-environment generalization.

Critical weaknesses

W1: PhysCorr relies heavily on the accuracy of physical states inferred by the vision-language model, yet their accuracy is not directly evaluated. It therefore remains unclear whether the reliability-conditioned intervention mechanism can effectively mitigate errors caused by inaccurate state estimates.

W2: The experimental validation is limited. The performance gain on VIAC is marginal, and EVIAC is an extension of the same dataset. The lack of evaluation on independent datasets limits the evidence supporting the effectiveness and generalizability of the proposed method.

Questions for authors

Q1: How is the accuracy of the physical states inferred by the vision-language model evaluated? How robust is the proposed correction mechanism when these state estimates are inaccurate?

Q2: What is the one-time cost of offline physical state generation, and what is the per-image inference latency of PhysCorr after caching the physical states?

Q3: Since the counterfactual construction modifies only the physical state while keeping the input images unchanged, how does this strategy promote invariant representations and improve cross-environment robustness?

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

Related papers

Discussion (0)

Sign in to comment.