acceptodds
Under review as a conference paper at ICLR 2027

Modeling Temporal Difference Cancellation for Remote Sensing Change Detection

Abstract

In binary remote sensing change detection, absolute feature differences discard sign patterns that influence how neighboring responses combine during spatial interpolation. Local differences with identical magnitudes can therefore produce different reconstructed responses. We propose a representation of temporal differences that accounts for this cancellation during spatial reconstruction. A shared hierarchical encoder and a shared projection produce signed differences between the two temporal feature maps. We compute a cancellation descriptor by subtracting the absolute value of the interpolated signed difference from the interpolated absolute difference. For interpolation weights that are nonnegative and sum to one, the descriptor is nonnegative, bounded above by the interpolated absolute difference, and invariant to exchanging the two input images. It captures local cancellation information that absolute differences alone do not encode. Lightweight residual readouts jointly process difference magnitude and cancellation at two reconstruction resolutions. Their outputs feed into the decoder alongside the original temporal feature pathway, allowing the decoder to learn how cancellation contributes to change prediction. The model is trained end to end with binary cross-entropy and Dice losses, requiring no additional task-specific annotations or auxiliary objectives.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Remote sensing, Binary change detection, Bi-temporal representation learning, Temporal difference cancellation, Feature interpolation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_28e4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This article begins with the observation that, in many recent studies on CD, part of the improvement is attributed to the fact that the proposed method combines features from the two images using descriptors (e.g. feature difference of multiplication) instead of letting the network learn these relationships itself. However, it is never clear whether the improvements actually stem from these interactions between the features of the two images alone, or if the operations used during decoding, in particular the type of fusion, also plays a role. The paper therefore studies whether the benefit of using descriptors depends on how feature are updated and fused during decoding. Using a toy model, they test the four scenarios: with and without descriptors and with joint or independent fusion of the features. The finding is that explicit descriptors help much more under joint refinement.

Critical strengths
  • The motivation of trying to untangle performance improvements is interesting and could be valuable

  • The design of the model and experiments, especially the two-factor experiment, as well as the evaluation protocol is clean

  • Limitations are reported clearly

Critical weaknesses
  • The main, and critical weakness is the limited generalization of the central claim. The article is meant as a general study of what the benefits of descriptors in CD model can be attributed to. Yet the comparison is only done on one architecture and one decoder family, and the central experiments, which is the factorial comparison, is only done on two datasets. The conclusion therefore only holds for this specific configuration and the global takeaway of the paper is therefore narrow. As it is, although very interesting, it sounds more like an ablation study of a given model. To make the paper much stronger and more convincing, you would need to have many recent strong CD methods, both in their BASE and JOINT version, and with and without joint training and make conclusion at a more global level.

  • No external comparison. The paper only compares with its own baselines that are ablations or small modifications of the same architecture, never with existing SOTA models. Again, this make the experiments look more like ablation studies of a broader study.

  • The components are not new. The descriptors mentioned have been used since early deep learning CD work, the architecture is not new. The paper shows that the joint fusion, which is the common way to decode the features, so nothing new, is the best practice. Even if this practice had maybe never been tested, this is a very modest contribution.

  • Given the fact that the paper focuses on how features are processed in the decoding stage, the paper would benefit from a stronger related work section on this part.

Things to improve the paper that did not impact the score:

  • Overall the paper is pretty difficult to follow. The paper would for example benefit from a much stronger introduction on what descriptors are and on the problematics that the paper is trying to address.
  • Some notations are not defined (K_3x3 for instance), and some are inconsistent, for example between section 3 and Figure 1.
  • BASE is mentioned early in the paper, yet define late. It is therefore hard to understand what it is when reading the paper.

Overall, the paper has a good research question and experimental design, and it is nice to underline that papers and CD models should investigate the origine of gains attributed to feature interaction, however the overall contribution of the paper is modest, and is closer to a careful ablation study. Only one architecture and two datasets are used for the main experiment that support the claim, yet the claim of the paper is global. The authors need to prove that the conclusions also hold for other architectures and other recent CD models.

Questions for authors
  • Is there an explanation or an intuition behind the fact that jointly refining the features works better than refining them independently?

  • The experiments are made with both descriptors at the same time. If you were to consider them separately and make the experiments one descriptor at a time, would the results be different? Does one descriptor benefits from the joint refinement more than the other?

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

This paper proposes JOINT, a change detection method which integrates joint residual refinement and multiscale reconstruction to improve the feature groups updates during multiscale decoding. Experiments across four datasets show modest improvements over a shared-encoder baseline.

Despite careful reading, I found the motivation and logical flow difficult to follow. It was not clear to me what the paper primarily aims to establish.

Critical strengths
  1. JOINT achieves improvements across four change detection benchmarks.
Critical weaknesses
  1. Unclear motivation. The paper does not clearly explain why the temporal descriptors should be investigated in relation to feature update strategies during multiscale decoding, nor does it explain why conventional feature fusion is insufficient. Also, the submission abstract appears not well aligned with the manuscript.

  2. Limited methodological novelty. The main contribution appears to be joint residual refinement of content and relation features, but the paper provides limited explanation of why this design improves feature learning.

Questions for authors
  1. Why is the feature state fixed at 192 channels? How is this state constructed, and what feature dimensions are used throughout the architecture given that 2w=1922w = 192? How sensitive is the performance to this choice, and does the proposed joint refinement generalize to other channel dimensions and models with different feature dimensions?
  2. Could the authors clarify the relationship between the method section titled Content–Relation Pyramid and JOINT? How do CC and QQ in Eq. (2) correspond to the content and relation stems in Figure 1, and what are their respective roles? In particular, why does the second convolution increase the channel dimension, and how does this operation contribute to joint refinement?
  3. What is the specific advantage of joint refinement over conventional feature fusion? Could the authors provide further analysis explaining how channel mixing and joint feature updates improve the utilization of temporal descriptors?
  4. Given the relatively modest F1 improvements over BASE, could the authors discuss the practical significance of these gains and provide further analysis of whether JOINT generalizes beyond the evaluated decoder configurations?
1 comment

My core concerns remain for this manuscript, especially regarding the tchnical contribution and the paper's motivation. Other reviews raised similar concerns. I thought the authors need to consider all these perspecitves to prepare a substantially improved version. So I maintain my origianl score.

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

This paper investigates the interaction between temporal feature construction and multiscale refinement strategies for remote sensing change detection. The authors propose JOINT, a content relation pyramid architecture that integrates temporal concatenation, absolute differences, and element-wise products through joint residual refinement, followed by multiscale aggregation and progressive spatial reconstruction. A two-factor experimental design is introduced to examine whether explicit temporal descriptors provide larger benefits under joint refinement than under independent refinement.

Critical strengths

The two-factor experimental design provides a systematic comparison of these design choices, going beyond conventional single-component ablations. The authors also conduct experiments on four benchmark datasets and report results across five random seeds. The evaluation includes multiple performance metrics, spatial error analysis, and computational costs, providing useful empirical observations about the proposed architecture.

Critical weaknesses
  1. Temporal feature interaction and multiscale fusion have already been extensively investigated in remote sensing change detection. The proposed investigation appears to address a relatively narrow architectural design choice or an ablation study rather than an important unresolved research problem. The authors should clarify why answering this question is scientifically meaningful.
  2. The main experimental section focuses almost exclusively on comparisons with the authors' baseline and structural variants. The quantitative comparisons with existing methods are presented only in Appendix I (Table 14).
  3. Although the paper discusses several recent methods published in 2026, including DIPFNet, AdaDINO, and EMDNet, these methods are not included in the quantitative comparisons. Appendix H even provides an architectural comparison with DIPFNet and AdaDINO, making the absence of corresponding performance comparisons particularly noticeable.
  4. The internal ablation experiments indicate that the proposed design provides relatively modest improvements. For example, JOINT improves F1 over WIDE-CONTENT by only 0.326 and 0.284 percentage points on LEVIR-CD and WHU-CD, respectively. However, the external comparisons show substantially larger advantages over certain existing methods.
  5. The paper does not present a sufficiently coherent progression from research motivation and problem formulation to methodological contributions and experimental conclusions. Several terms, such as "explicit temporal descriptors," are introduced without adequate conceptual explanation. The notation is also confusing in places. For example, (C_i) in Figure 1 represents the upsampled coarse-scale feature, whereas the same notation is used for the content stem in Equation (2).
  6. The proposed method improves F1 by approximately 0.5–0.66 percentage points over BASE, but substantially increases inference time and memory consumption. Under AMP with batch size eight, inference time increases by approximately 38%, while peak memory consumption increases from 340 MiB to 1,003 MiB. The paper does not sufficiently justify whether these additional computational costs are warranted by the relatively modest improvements.
Questions for authors
  1. What is the fundamental research problem addressed by this paper, and what important new insight does it provide beyond existing studies of temporal feature interaction and fusion?
  2. Why are the main quantitative comparisons with existing methods presented only in the appendix? Could the authors incorporate these results into the main paper and provide a more comprehensive analysis?
  3. Why are recent and closely related methods, such as DIPFNet and AdaDINO, not included in the quantitative evaluation? If direct comparison is inappropriate, what are the key differences in their research settings or objectives?
  4. The ablation experiments show relatively modest improvements from the proposed design, whereas the external comparisons show much larger advantages over some methods. Which architectural components primarily account for these differences? Could the authors provide more controlled evidence to identify the key contributing factors?
  5. Could the authors improve the clarity of the research motivation, terminology, and notation, and explicitly summarize the main scientific findings and their implications?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.