- 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
-
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.
-
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
-
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.
-
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.
-
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
- 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?
- 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?
- 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?
- 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?
- 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?