acceptodds
Under review as a conference paper at ICLR 2027

BiRNeT: A Unified Bidirectional Reasoning Paradigm for Open-Vocabulary Remote Sensing Change Detection

Abstract

Open-vocabulary change detection (OVCD) aims to identify changes of arbitrary categories based on text descriptions, overcoming the reliance of conventional methods on predefined categories. However, existing approaches generally treat change localization and semantic recognition as separate processes, and the lack of interaction between them has become a key bottleneck for further performance improvement. In this paper, we propose BiRNet, a unified bidirectional reasoning paradigm. It uses change probabilities to constrain bi-temporal semantic information and performs joint inference to achieve unified prediction of change states and change categories. To support this paradigm, an Adaptive Multi-scale Context-Difference Fusion Module (AMCDF) is designed to enhance change representation. A Bidirectional Change-Constrained Semantic Transition Module (BCCST) is further proposed to couple localization and recognition into a joint decision through a semantic transition matrix constrained by change probabilities. BCCST requires no additional training and offers plug-and-play capability. Experimental results on four benchmark datasets demonstrate that BiRNet achieves competitive performance, with significant F1 gains of 23.69% and 6.93% on S2Looking and SECOND, respectively. The code will be released upon publication.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Open-vocabulary change detection, Remote sensing

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_18NpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies open-vocabulary change detection in remote sensing. It argues that pipelines such as Seg2Change handle change localization and semantic recognition separately, and that this lack of interaction limits performance. BiRNet has a binary change branch (frozen SAM ViT-B encoder plus a trained fusion module, AMCDF) and an off-the-shelf open-vocabulary segmenter (CAFe-DINO). A training-free module, BCCST, builds a K x K transition matrix from the two per-date class distributions, reweights the diagonal by 1-q and the off-diagonal by the change probability q, and takes the arg-max to decide both whether a pixel changed and its from-to classes. Results are reported on LEVIR-CD, WHU-CD, S2Looking and SECOND against six recent methods.

Critical strengths
  1. BCCST is simple and reusable. It replaces hard masking of the semantic maps with a probabilistic decision, has no parameters, and can be attached to any binary change detector and any open-vocabulary segmenter. I think this is the most useful part of the paper.

  2. The comparison set is up to date. Four standard datasets and six recent OVCD methods are covered, and the paper also reports the cases where it is not the best, such as WHU-CD.

Critical weaknesses
  1. The gains do not come from the bidirectional reasoning that the paper is built around. BCCST is called the core module, but in Table 4 adding it changes IoU by +2.16 on LEVIR-CD, -0.22 on WHU-CD and +1.26 on S2Looking, and mIoU on SECOND by +0.18. The Baseline row, with neither AMCDF nor BCCST, already reaches 57.79 F1 on S2Looking and 52.53 mF1 on SECOND, while the best previous method has 39.57 and 46.76. So about 18 of the 23.69 F1 points claimed in the abstract for S2Looking are there before either proposed module is added.

  2. Table 3 points the same way. BCCST adds about 2 to 3 IoU to Seg2Change* on three datasets and slightly hurts on S2Looking, and the result stays far below the original Seg2Change (46.35 vs 64.89 IoU on LEVIR-CD). Section 4.4 reads this table as evidence that the improvement "mainly comes from the effective interaction", which I do not see in the numbers.

  3. The training data of the change branch is never stated. Sections 3.5 and 4.2 say this branch is trained for 50 epochs, and Section 4.1 lists only test splits. If it is trained on the training split of each benchmark, then Tables 1 and 2 compare an in-domain supervised detector with baselines that include training-free pipelines such as the DynamicEarth variants. That would also explain why the Baseline alone beats all prior work on S2Looking by such a margin.

Minor points:

  • Notation and names drift. The superscript is IoU^c in Tables 1-2 and IoU^C elsewhere and is never defined, the SECOND class is "surface" in Section 4.1 but "bareland" in the Fig. 4 caption, and the SECOND mIoU of the full model is 38.29 in Table 2 and 38.28 in Table 4.
  • Differences in points are written as percentages, e.g. "gains of 23.69% and 6.93%" in the abstract and "6.8% and 6.93%" in Section 4.3.2.
  • The column labels in Fig. 4 are too small to read.
Questions for authors
  1. What data is the change branch trained on? Is it one model for all four benchmarks or one per dataset, and does the training data include the training splits of the evaluation datasets?
  2. Please define the "Baseline" and "w/o BCCST" rows in Table 4. How is the semantic change map produced without BCCST, and what replaces AMCDF in the Baseline?
  3. With the same segmenter, Seg2Change* gets 44.37 IoU on LEVIR-CD while your model without BCCST gets 73.69. Where does this gap come from if not from the change branch and its training data?
  4. Eq. 9 normalises the K diagonal entries and the K^2-K off-diagonal entries separately. With many classes the largest off-diagonal entry may become small relative to the largest diagonal one, which would push decisions toward "no change". How does the rule behave as K varies, and how many classes are prompted on the building datasets?
  5. The fusion in Eq. 6-7 (a local and a global branch on the summed features, a sigmoid weight w, then a weighted sum of the two inputs with weights w and 1-w) looks very close to attentional feature fusion (Dai et al., WACV 2021), which is not cited. How does AMCDF differ from it?
Reviewer_hZQbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses open-vocabulary remote sensing change detection and argues that existing pipelines treat change localization and semantic recognition as independent processes. BiRNet trains a binary change branch on top of a frozen SAM encoder with an Adaptive Multi-scale Context-Difference Fusion module, obtains bi-temporal semantic probabilities from a pretrained open-vocabulary segmentation model, and combines the two through a training-free Bidirectional Change-Constrained Semantic Transition module. BCCST builds a pixel-wise semantic transition matrix, constrains it with the change probability, and jointly decides the change state and the categories before and after the change. Experiments are conducted on LEVIR-CD, WHU-CD, S2Looking, and SECOND.

Critical strengths
  1. The paper identifies the lack of interaction between change localization and semantic recognition as a gap in open-vocabulary change detection, and BCCST offers a simple, parameter-free way to couple the two probability outputs. When plugged into Seg2Change*, it improves results on three of the four datasets. Its simplicity and plug-in nature make it practically useful.

  2. The overall system achieves strong results on four benchmarks, with large F1 gains on S2Looking and SECOND.

Critical weaknesses
  1. The source of the overall improvement does not match the paper's claim. The abstract and Sec. 4.4 attribute the improvement mainly to the interaction between change and semantic information. However, in Table 4 the Baseline already reaches 57.79 F1 on S2Looking and 52.53 mF1 on SECOND, i.e., 18.22 and 5.77 points above the best compared method. BCCST itself changes F1 by +1.42, −0.14, +1.19, and +0.12 on the four datasets. In Table 3, Seg2Change*+BCCST remains below the original Seg2Change on all four datasets, which does not support the conclusion in Sec. 4.4 that the gain comes from the interaction rather than from a stronger semantic model.

  2. The supervision and comparison protocol is unclear. Sec. 3.5 states that the change branch is trained with binary change labels, but does not specify the training data or whether it is trained separately for each benchmark, whereas several compared methods (e.g., AdaptOVCD, OpenDPR, and OmniOVCD) do not train on target-domain labels. If the change branch is trained on each benchmark's training split, the large leads in Tables 1–2 would reflect in-domain supervision rather than an advantage under a shared open-vocabulary setting.

Questions for authors
  1. On which data is the change branch trained for each benchmark, and what are the supervision conditions of each compared method in Tables 1–2?

  2. What exactly is the Baseline in Table 4, and how are the change and semantic outputs combined in the "w/o BCCST" variant?

  3. After the normalization in Eq. 9, the diagonal and off-diagonal entries sum to approximately 1−q and q, respectively. The decision therefore appears equivalent to thresholding the change probability with a threshold that depends on how concentrated the semantic evidence is, while the magnitude of the heterogeneous evidence E_diff itself no longer affects the decision. Could you comment on this, and compare BCCST with simple semantic filtering or probability fusion?

  4. Could you report the accuracy of from–to transition pairs on SECOND (i.e., both the pre-change and post-change categories correct), which would more directly support the joint reasoning claim?

Reviewer_R294OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BiRNet for open-vocabulary remote sensing change detection. It combines a binary change detector with an open-vocabulary segmentation model through two modules: AMCDF fuses contextual and difference features across scales, while BCCST uses change probabilities to guide semantic transition predictions. Experiments on several datasets show improvements over existing methods.

Critical strengths
  1. The paper addresses a clear problem, change detection and semantic segmentation can produce inconsistent predictions.

  2. The results show a substantial improvement. The experiments cover both building changes and multi-class land-cover changes.

  3. The paper includes component ablations, helping assess the contributions of feature fusion and joint prediction separately.

Critical weaknesses
  1. The ablations show limited gains from BCCST. The paper should clarify this baseline and compare simpler fusion rules under the same training setup.

  2. The experiments do not directly test generalization to unseen categories. Three datasets focus on buildings, and SECOND uses six fixed classes. The paper should specify the category splits, text prompts, and supervision used, and evaluate performance when new categories are introduced.

  3. The method treats matching semantic labels as no change, which excludes changes within the same category. This limitation needs discussion. The paper also lacks run-to-run variation and runtime and memory measurements, making the small BCCST gains and computational cost harder to assess.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.