acceptodds
Withdrawn from ICLR 2027

Resolution-Aligned Change Detection via Diffusion Super-Resolution and RL-Guided Unseen-Class Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-resolution change detection (CD) is challenging because differences in spatial resolution between temporal observations introduce structural misalignment, inaccurate boundaries, and false change responses. An additional challenge arises when semantic classes excluded from base-detector training may appear after deployment without target-scene pixel-level annotations. We introduce a resolution-aligned adaptation framework that combines conditional diffusion-based super-resolution with reinforcement-learning-guided pseudo-label refinement. First, a conditional diffusion model reconstructs the low-resolution after image in the spatial domain of the high-resolution before image using noise-prediction, reconstruction, and perceptual objectives that preserve geometry and empirically reduce reconstruction-induced artifacts. A shared-weight Siamese detector with cross-temporal attention then predicts changes from the aligned image pair. To adapt beyond the original label space, candidate-mask refinement is formulated as a finite-horizon decision process. An RL policy iteratively refines, accepts, or rejects candidate regions using a retrieval-augmented reward combining geometric consistency, image-text alignment, and language-guided semantic feedback. Accepted masks provide pseudo-labels for target-mask-free adaptation to semantic classes held out from base-detector training. Experiments on synthetic and real cross-resolution benchmarks demonstrate improved reconstruction fidelity, change-detection accuracy, held-out-class coverage, mask quality, and post-adaptation detector performance over matched baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Change detection, remote sensing, cross-resolution learning, super-resolution, diffusion models, reinforcement learning, retrieval-augmented learning, semantic segmentation

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

This paper addresses cross-resolution remote sensing change detection and adaptation to semantic classes held out during base-detector training. The proposed pipeline first reconstructs the post-event image using conditional diffusion super-resolution, then performs cross-temporal change detection. A frozen segmentation model generates candidate regions, which a PPO policy iteratively refines, accepts, or rejects using retrieval-augmented multimodal rewards. An external semantic model assigns pseudo-labels to accepted regions, and only the additional class head is trained. The authors claim improvements in cross-resolution change detection and semantic coverage without acquiring manual pixel-level annotations for the new classes in target scenes, while largely preserving performance on previously seen classes.

Critical strengths
  1. The application problems are meaningful, and several task boundaries are clearly specified. Resolution mismatch and semantic expansion after deployment are practically relevant challenges. The paper distinguishes resolution alignment from geometric registration and clarifies the role of external semantic knowledge, making the setting more precise than a broad claim of zero-shot discovery.
  2. The evaluation is broad, including component ablations, manual verification, and performance retention on seen classes.
Critical weaknesses
  1. The central contribution lacks focus. The pipeline combines several established components, but the connection between resolution alignment and adaptation to new classes, and the need for a unified framework, remain unclear.
  2. The comparison protocols are inconsistent. Table 2 states that baseline results come from different publications and may use different protocols, whereas the text claims identical inputs, splits, and training conditions.
  3. Recent baseline coverage is limited. Table 2 contains no 2026 methods. Please justify the selection and discuss task-comparable work available before manuscript completion.
  4. SR training conditions need clarification. Were SwinIR and OSEDiff in Table 2 fine-tuned on the same remote sensing data, and which checkpoints and budgets were used? Matching output dimensions and the detector does not control for domain-adaptation differences.
  5. The independent benefit of RL is insufficiently established. A greedy multi-step baseline with the same reward and model-call budget is missing. Acceptance precision/FDR alone cannot exclude improvements obtained by accepting fewer candidates.
Questions for authors
  1. Which baselines in Table 2 were rerun under matched conditions? Were the SR baselines in Table 4 fine-tuned on the same remote sensing data?
  2. What ground-truth labels and denominators were used for held-out-class coverage and post-adaptation F1/mIoU?
  3. Do you have results comparing RL with greedy refinement under the same reward and computational budget, including both precision and recall?
Reviewer_UDAaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies remote-sensing change detection under resolution mismatch and its adaptation to semantic classes excluded from the task-specific training set. The framework first reconstructs the low-resolution image using conditional diffusion super-resolution, then it applies a Siamese change detector with cross-temporal attention. For adaptation, it generates candidate regions using semantic segmentation differences, and a PPO policy iteratively refines, accepts, or rejects them using geometric, image-text, and retrieval-augmented language-model rewards. For those accepted masks,it receives semantic labels from an external frozen model and supervise a new classification head. Experiments cover three cross-resolution benchmarks, component ablations, and human verification of accepted pseudo-labels.

Critical strengths
  1. Resolution mismatch and incomplete semantic coverage are meaningful challenges in remote sensing. Combining the reconstruction with subsequent detector adaptation provides useful insights.
  2. The paper attempts to evaluate multiple stages of the pipeline. Table 4 compares several resolution-alignment methods using a common detector, while Tables 3 and 5 examine adaptation components. The human review of 1,000 accepted masks and the seen-class retention analysis are also useful. These experiments provide a foundation for assessing the system, although several protocol and interpretation issues currently limit their conclusiveness.
Critical weaknesses
  1. Resolution mismatch and incomplete semantic coverage are meaningful challenges in remote sensing. Combining the reconstruction with subsequent detector adaptation provides useful insights.
  2. The paper attempts to evaluate multiple stages of the pipeline. Table 4 compares several resolution-alignment methods using a common detector, while Tables 3 and 5 examine adaptation components. The human review of 1,000 accepted masks and the seen-class retention analysis are also useful. These experiments provide a foundation for assessing the system, although several protocol and interpretation issues currently limit their conclusiveness.

Critical Weaknesses 1、Table 2 states that baseline results are reproduced from published work. However, the paragraph below claims that all methods use the same inputs, splits, degradation factors, and training conditions. The authors should identify which results were obtained through controlled reruns and which were taken from published work. 2、Lines 455–457 state that no native HR after-image reference is available, which is why the HR-after oracle is omitted. Nevertheless, Table 1 reports reference-based PSNR and SSIM for DE-CD, including 37.45 dB and 92.55 at the 64×64 input setting. What reference images were used? If these results come from a separate synthetic reconstruction experiment, they must be distinguished from evaluation on the actual cross-sensor inputs. 3、In the cited SILI paper, the DE-CD construction described uses Sentinel-2 pre-event images and Planet post-event images, whereas this submission assumes HR-before/LR-after inputs. Reversing pairs may be reasonable, but the transformation and its implications for semantic after-image labels should be documented before comparing published results. 4、This work defines its study as target-mask-free rather than category-agnostic, and allows the external semantic labeler to contain knowledge of held-out categories. My concern is therefore not the mere use of external semantic knowledge, but how performance under this setting is measured and attributed. 5、Table 3 reports the Coverage, Class Accuracy, and Mask IoU are evaluated using an independent frozen model. Separating this evaluator from pseudo-label generation is useful, but agreement with another model is not equivalent to agreement with ground truth. The paper does not sufficiently establish the evaluator’s accuracy on the relevant classes and sensors, or provide precise matching rules and denominators for these metrics.

6、The human assessment of 1,000 accepted masks supports the reported 95.9% acceptance precision. However, reviewing accepted predictions alone cannot establish recall, coverage, or dense boundary accuracy. A separately annotated evaluation subset containing both recovered and missed changes is needed to substantiate these claims.

7、There is also an unresolved question about the base-detector row in Table 3. The base semantic head is described as predicting only seen classes, yet the table reports 65.2% Class Accuracy in an evaluation presented as held-out-class adaptation. Please clarify whether this includes seen classes, uses an external semantic labeler for the base detector, or measures something else. Likewise, specify whether the adapted F1 and mIoU are binary or semantic metrics, and whether adaptation and final testing use disjoint images. 8、The paper compared several static-filtering and reward-component baselines. However, these comparisons primarily establish an advantage over one-shot decisions. They do not distinguish the benefit of PPO from the benefit of performing multiple mask edits and repeated reward-model evaluations. A suggested baseline:the same expand, contract, and shift operations through greedy search or another simple iterative procedure, using the same quality function, horizon, retrieval pool, and reward-query budget. This would test whether learning a sequential policy provides additional value over directly optimizing the available score. 9、Table 5 reports acceptance precision and FDR, not reporting acceptance counts or matched-coverage comparisons. Although Table 3 provides evaluator-based coverage for some variants, the full ablation should report accepted-mask counts, precision–coverage trade-offs, and downstream performance under comparable pseudo-label budgets. 10、 Other comments The discussion of LEVIR-CD treats unannotated roads, vegetation, and swimming pools as annotation failures. LEVIR-CD explicitly targets building changes; detecting other categories extends the task rather than automatically correcting its ground truth. Results on the expanded semantic task should be separated from standard building-change evaluation. • Figure 9’s caption states that AUC peaks near an IoU threshold of 0.18, but the displayed curve continues increasing substantially beyond that point. The distinction between selecting an operating threshold using Youden’s J and computing ROC-AUC over a score distribution also needs clarification. • Please provide per-dataset and per-held-out-class adaptation results, together with variability across runs or evaluation folds. The current aggregate results conceal potentially important differences between frequent and rare categories.

Questions for authors
  1. Which baselines in Table 2 were run under identical conditions, and what reference images were used to compute DE-CD PSNR and SSIM?
  2. What are the exact definitions, reference labels, and aggregation rules for the Table 3 metrics? How can a seen-class-only base head obtain nonzero held-out-class classification accuracy?
  3. Does PPO outperform an iterative greedy-refinement baseline with the same actions, reward function, and query budget?
  4. Please reconcile the taxonomy and implementation inconsistencies, and correct the convergence claim in Appendix L.1?
Reviewer_cP2bOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines diffusion-based super-resolution with change detection for satellite image pairs with different resolutions. It then uses retrieval and an RL policy to refine candidate change masks and generate pseudo-labels for adapting to held-out classes. Experiments cover three benchmarks, with ablations of the SR and adaptation stages.

Critical strengths
  • The paper addresses a practical problem: detecting changes when image resolution differs and some classes lack target-scene annotations.

  • The SR comparisons and adaptation ablations are useful. The human review of 1,000 accepted masks also provides evidence that the selected pseudo-labels are reasonably reliable.

Critical weaknesses
  • The comparison protocol is inconsistent. Table 2 says baseline results come from published papers and may use different evaluation protocols, while the text below says all methods use the same inputs, splits, and training conditions (p. 8). This makes the claimed advantage difficult to assess.

  • The evaluation needs clearer references. Table 1 reports PSNR/SSIM on DE-CD, but p. 9 says no native HR after-image reference is available. Also, the held-out-class metrics rely on a separate model whose accuracy against human annotations is not established. These gaps make it hard to judge the reconstruction and adaptation gains.

Questions for authors
  • Which Table 2 results were rerun under your protocol, and which were copied from prior papers?

  • What reference images were used to compute DE-CD PSNR/SSIM, given the lack of native HR after images?

  • How are Coverage, Class Accuracy, and Mask IoU defined, and how was the independent evaluator validated against human annotations?

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

Related papers

Discussion (0)

Sign in to comment.