acceptodds
Under review as a conference paper at ICLR 2027

RIFT-CD: Preserving Spatial Detail and Relational Invariance for Robust Remote Sensing Change Detection

Abstract

Remote sensing change detection requires simultaneously preserving fine-grained spatial evidence and learning temporal representations that generalize across diverse scenes. However, modern pretrained vision backbones often compress spatial information aggressively, while conventional temporal fusion can still exploit scenespecific appearance cues, limiting robustness under small structural changes and distribution shifts. We introduce RIFT-CD, a relationally normalized, resolutionpreserving backbone for change detection. Rather than treating temporal reasoning as an auxiliary prediction component, RIFT-CD integrates change-sensitive representation learning directly into the feature trunk. The architecture maintains high-resolution semantic representations throughout the hierarchy, adopts information-preserving feature transitions, and performs temporal interaction in a content-normalized relational space. We further regularize backbone adaptation to retain transferable pretrained representations while preventing late-stage specialization to training scenes. This design provides a unified way to balance spatial fidelity, temporal discrimination, and cross-scene generalization within a single end-to-end model. Experiments across multiple change-detection benchmarks demonstrate consistent improvements in segmentation quality, boundary localization, and robustness to appearance variation, suggesting that preserving information and constraining representation drift are central to building more generalizable change detectors.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Change Detection, Representation Learning, High-Resolution Vision, Temporal Representation Learning, Domain Generalization, Pretrained Vision Backbones

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_YjcYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper identifies 2 limitations of current CD architectures, namely 1/ the fact that pretrained vision backbones compress too much the information, leading to difficulties when it comes to detecting tiny changes; and 2/ the fact that current models tend to cross information between the images already during encoding through the interaction of features extracted from the two images, and while this can help carry structural changes it can also carry irrelevant variations like illumination or seasonal differences. The paper makes 2 main contributions: First, they make a study of how and when does crossing information between the two images help. The study shows that the benefits depend on the information being exchanged: with raw differences and pair-mean appearance, cross information during the encoding stage hurts performance, while normalization helps stabilizing the information agains noise, and replacing the pair mean by channel-wise agreement increases change selectivity. Second, they introduce RIFT-CD, a CD model build on an adapted DINOv3 model that uses these results and preserves high-resolution features to answer the two limitations above, achieving SOTA performance.

Critical strengths
  • The main comparison/ablation for cross-temporal feedback is interesting. In particular the factorial design (Table 3) is valuable.
  • The “S-R-F” (=Stride-2 stem instead of Stride-4 to increase resolution, Normalized relation and temporal feedback) decomposition is clear.
  • Honest reporting, including the costs that come with every update (e.g. 3.75x compute when using a Stride-2 stem instead of Stride-4, etc)
  • Tested robustness against radiometric perturbations
  • Measuring message drift and update selectivity help understanding, beyond general F1 performance, the impact of each component.
Critical weaknesses
  • Modest novelty: It seems like every ingredient (feature interaction in the encoder à la Changer, backbone, descriptors) already exists in the literature. The contribution is the combinaition of those ingredients and the associated justification. As underlined in the contributions listed in the introduction, the main contribution of the paper is the study of the importance of the crossed information/feedback. The RIFD-CD architecture is merely the resulting architecture that arise from the conclusion of this study, plus some additional tweaking of the backbone strides to increase the resolution. The model, as it is, does therefore not bring much novelty. The study, however, does! But unfortunately, the latter is reduced to a single dataset, LEVIR-CD, and a single architecture. What happens for example with other pretrained encoders? Other types of CD architectures? Other images from other datasets with different observable objects/scenes and different resolution? The paper could be very interesting if more general and generalizable, yet all the interesting results are only provided on a single dataset with a single architecture. The claim is general, but not proven, and this is a major weakness of this work.

  • Robustness is only tested with synthetic perturbations. What about real distributions shifts?

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

  • The notation in Figure 1 can be a bit confusing. You denote S,R,F the different factors but S is for example already used for the outputs of the RC module. Similarly, the colors are not consistent, S is in blue and so are the features of the first image, but F is in orange like the features of the second one. It could be nice to have a clear color code to make the different factor visually clear. You could also add Ht, Ft, al, bl or Xl in the figure to help the reader understand the methodology.
  • Where is the normalization applied? It is not clear in the paper. Is it before computing the descriptors?
  • Some notations could be better introduced, as some are never defined. What is m_l in line 375 for example? The definition is, if I am not wrong, only in the supplementary materials.

General assessment: The paper asks a great question which is when and how does feature interaction help change detection in remote sensing. The paper is rigorous and the comparisons (factorial design) are well constructed and interesting, yet the scope of this work is narrow. The central findings come from only one dataset, one architecture, one interaction mechanism, despite the claim being general. The only thing that is demonstrated is that the conclusion holds in this configuration. And without this stronger analysis, the resulting RIFT-CD seem incremental. The architecture can be understood as an output to the study, rather than an independent contribution (which is fine!!) but in that case the study needs to be stronger and more general.

Questions for authors

See Weaknesses

2 comments

Critical Weaknesses

CW1. Novelty and generality of the mechanism study. We agree that the individual components of RIFT-CD are not claimed as fundamentally new architectural primitives. Rather, the central contribution of this work is a controlled investigation of how relational evidence affects the utility of cross-branch temporal feedback. Specifically, our factorial design isolates the interaction between evidence completeness (Raw vs. Full) and feedback direction (Self vs. Cross), allowing us to examine whether relational evidence changes the relative effectiveness of cross-branch feedback beyond what can be achieved through self-context feedback alone. RIFT-CD is the resulting architectural instantiation of these findings, rather than the sole source of novelty.We have expanded the study along all three dimensions you requested.

1. Other pretrained encoders. LEVIR-CD held-out test F1 (%):

EncoderFull-SelfFull-CrossGain (pp)
ConvNeXt-T91.64691.697+0.051
Swin-T91.67191.729+0.059
ConvNeXt-V2-T/FCMAE91.83291.877+0.045

Full-Cross improves over Full-Self on all three added encoders. The Full-versus-Raw feedback interaction, (Full-Cross − Full-Self) − (Raw-Cross − Raw-Self), is positive for ConvNeXt-T, Swin-T and ConvNeXt-V2-T: +0.079/+0.044/+0.255 F1 pp, respectively. For ConvNeXt-V2-T, enabling feedback also improves over the feedback-disabled control by +0.642 F1, +1.092 IoU and +0.721 Boundary-F1 pp; Swin-T improves all three metrics as well.

2. Other CD architectures and datasets. We integrated the normalized directed-feedback adapter into Siamese U-Net and Changer. Held-out test comparisons against each original architecture are:

Dataset / architectureOriginal F1Cross F1ΔF1 / ΔIoU (pp)
LEVIR / Siamese U-Net91.20991.268+0.059 / +0.100
LEVIR / Changer90.45290.944+0.492 / +0.823
MineNetCD256 / Siamese U-Net64.95466.710+1.755 / +1.950
MineNetCD256 / Changer63.74365.801+2.057 / +2.250

On MineNet, Siamese U-Net Cross also exceeds the strong self-context control by +1.201 F1 and +1.340 IoU pp. These results demonstrate portability to two additional CD architectures and a new dataset.

3. Replicated mechanism evidence. On S2Looking, Full-Cross improves over Full-Self by +0.288 ± 0.337 F1 and +0.310 ± 0.237 Band-F1 pp (paired mean ± sample SD). S2Looking adds different scenes, viewing geometry and spatial resolution. Full is the only evidence arm with positive mean feedback gains in both metrics. Full also reduces radiometric/noise message drift at all three feedback levels in every seed on both datasets.

The additional SYSU-CD test gives Full-Cross 86.074 F1, 75.553 IoU and 64.323 Boundary-F1, improving over Full-Self by +0.057/+0.087/+0.285 pp. The expanded evidence connects the mechanism analysis to gains across encoders, architectures and datasets.

More importantly, the SYSU-CD factorial comparison shows that Raw-Cross decreases F1 by 0.1667 pp relative to Raw-Self, whereas Full-Cross improves F1 by 0.0567 pp. This yields a positive interaction of +0.2234 F1 and +0.3431 IoU pp across three seeds, further supporting the importance of complete relational evidence for cross-branch feedback.

CW2. Real distribution shifts. We added complementary evaluations on real imagery. Source-only cross-dataset transfer. Models use source-validation checkpoints and threshold 0.5, with no target fine-tuning, checkpoint selection or threshold tuning. F1 (%) is mean over three seeds:

Source → targetRIFTChanger
LEVIR → S2Looking40.4419.64
LEVIR → WHU85.5680.24
S2Looking→LEVIR78.8935.11

These evaluations extend the robustness evidence to naturally occurring appearance changes and unseen target datasets.

Things to improve the paper that did not impact the score

Notation and colors. The symmetric change state is TlT_l; S,R,F\mathsf S,\mathsf R,\mathsf F denote design factors. Neutral factor tags and consistent blue/orange temporal branches distinguish these roles; labels Xlt,al,bl,Ht,FtX_l^t,a_l,b_l,H^t,F^t connect the figures to the equations.

Normalization order. Channel-wise L2 normalization N(x)=x/max⁡(∣x∣2,10−6)\mathcal N(x)=x/\max(\\|x\\|_2,10^{-6}) is applied at each spatial location to projected features and attention outputs before constructing Dl=∣aˉl−bˉl∣D_l=|\bar a_l-\bar b_l|, Cl=∣cˉlA−cˉlB∣C_l=|\bar c_l^A-\bar c_l^B| and Al=aˉl⊙bˉlA_l=\bar a_l\odot\bar b_l.

Message definition. The main text defines mlA=Ψl([aˉl,cˉlA,Dl])m_l^A=\Psi_l([\bar a_l,\bar c_l^A,D_l]), ulA=λl⊙mlAu_l^A=\lambda_l\odot m_l^A and X~lA=XlA+ulA\widetilde X_l^A=X_l^A+u_l^A at first use; branch B is symmetric.

Dear Authors,

I acknowledge your rebuttal. Unfortunately, since it was posted shortly before the end of the discussion period, I may not have time to carefully review your responses and provide further constructive feedback before the end of the discussion period.

I will, however, consider your responses during the reviewers/AC discussion.

Best regards,

Reviewer_bZH4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RIFT-CD, a remote-sensing change detection framework built on a fully fine-tuned DINOv3 ConvNeXt-Large backbone. The method aims to jointly preserve fine spatial details and prevent cross-temporal feature interaction from propagating scene-specific appearance variation.

RIFT-CD contains three main design choices. First, it changes the pretrained 4×44\times4 stem from stride four to stride two, producing denser semantic feature maps throughout the encoder. Second, it constructs normalized relational descriptors from direct feature discrepancy, cross-attended contextual discrepancy, and channel-wise agreement. Third, it injects the resulting directed relational messages back into the two temporal branches before subsequent encoder stages. Native half- and full-resolution image differences are additionally used during decoding.

The method is evaluated on seven remote-sensing change detection datasets and reports the highest F1 score among the compared methods on all seven. The paper further presents spatial-resolution controls, an eight-model normalization–descriptor–feedback factorial study on LEVIR-CD, paired-scene bootstrap intervals, perturbation experiments, tiny-component recall, and feature-level probes intended to explain why normalized relational evidence makes cross-temporal feedback beneficial.

Critical strengths

(i) The paper provides a more mechanism-oriented analysis than is typical for change detection architectures. The N×M×FN\times M\times F study does more than report a conventional module ablation: it explicitly examines how normalization and pair-mean replacement alter the effect of the same feedback route. The message-drift, update-selectivity, perturbation, and feature-routing probes form a reasonably coherent explanation of the observed behavior.

(ii) The method demonstrates strong reported accuracy and particularly promising boundary localization. RIFT-CD achieves the highest reported F1 across all seven benchmarks. Under the matched LEVIR-CD control, it improves over SimpleDiff-S2 by 0.210.21 F1 points and 2.682.68 Band-F1 points. The comparatively larger boundary improvement is consistent with the stated motivation of preserving fine spatial evidence.

Critical weaknesses

(i) The comparison with ChangeDINO(strong competition method) does not isolate the architectural contribution of RIFT-CD. ChangeDINO uses a frozen DINOv3 feature extractor, whereas RIFT-CD fully fine-tunes its DINOv3 ConvNeXt-Large backbone. Therefore, the reported improvements may partly result from the different backbone adaptation strategies rather than the proposed relational coupling and cross-temporal feedback. A controlled comparison under the same frozen or fine-tuned DINOv3 setting is needed.

(ii) The benchmark comparison does not sufficiently isolate the contribution from model scale, pretraining, and computation. RIFT-CD uses a fully fine-tuned 201M-parameter DINOv3 ConvNeXt-Large model and approximately 399G counted operations per 256×256256\times256 image pair. The paper does not clearly state whether all Table 1 baselines were reimplemented under the same splits, augmentations, crop size, training budget, and checkpoint-selection policy, or collected from heterogeneous sources. The only tightly matched baseline is the internally defined SimpleDiff-S2 model. In addition, switching from stride four to stride two increases counted operations by 3.753.75\times$$ for a 0.320.32-point F1 and 2.962.96-point Band-F1 improvement. Without a matched strong interaction baseline—such as a Changer-style model using the same DINOv3-L backbone—and practical memory/throughput measurements, it is difficult to determine how much of the reported benchmark advantage comes specifically from the proposed relational feedback design rather than the unusually strong and computationally expensive backbone configuration.

Questions for authors

(i) How much of the improvement over ChangeDINO comes from fully fine-tuning the DINOv3 backbone? Could the authors compare RIFT-CD and ChangeDINO under the same backbone adaptation setting, either by freezing DINOv3 in both methods or fine-tuning it in both methods?

(ii) Are the baseline numbers in Table 1 obtained from the authors’ own reimplementations or copied from the respective papers? Please provide the exact training/evaluation protocol and checkpoint-selection rule for each benchmark.

2 comments

Critical Weaknesses

CW1. Separating architecture from backbone adaptation.

RIFT's advantage persists when both models use the same backbone and adaptation strategy. We evaluated controlled ChangeDINO-ConvNeXtL using identical DINOv3 ConvNeXt-L/LVD1689M initialization, retaining ChangeDINO's FPN, fusion, decoder, and morphology components. Frozen controls freeze only DINO; task modules remain trainable.

LEVIR-CD held-out test differences, separate from original validation, are RIFT minus baseline: paired mean ± sample SD in percentage points (pp):

AdaptationReported seed pairsΔF1ΔBand-F1
Frozen3+0.643 ± 0.094+1.890 ± 0.291
Fully fine-tuned3+0.427 ± 0.017+1.384 ± 0.063

Every reported pair favors RIFT under both adaptation settings.

CW2. Isolating interaction design and measuring its cost.

The requested Changer control shares the pretrained backbone and training protocol while retaining spatial/channel exchange, FDAF, and MixFFN. The backbone initialization, input resolution, decoder/head, loss, augmentation, optimizer, training budget, and checkpoint-selection policy are matched. The interaction design is the primary architectural difference.

All Changer-ConvNeXtL, SimpleDiff-S2, and RIFT-S2 models use the same fully fine-tuned DINOv3 ConvNeXt-Large backbone, initialized from the same LVD1689M checkpoint. All backbone stages are trainable under the same backbone learning rate.

Three-seed test comparisons are:

RIFT-S2 minusΔF1 (pp)ΔBand-F1 (pp)
Changer-ConvNeXtL+0.464 ± 0.104+1.303 ± 0.357
SimpleDiff-S2+0.119 ± 0.047+0.197 ± 0.086
RIFT-S4+0.390 ± 0.076+0.945 ± 0.064

Both metrics improve in every seed. SimpleDiff-S2 retains dense DINO semantics and native-detail paths, isolating the joint gain from relational coupling and feedback.

To isolate feedback direction, we also tested Raw/Full × Self/Cross on SYSU-CD with the same DINOv3-L backbone (4,000 test pairs). Cross−Self F1 changes from −0.1667 pp with Raw evidence to +0.0567 pp with Full evidence: an interaction of +0.2234 pp.

Practical costs on an exclusive RTX 4090 are:

Model, FP32 batch 8p50/p95 latency (ms)Pairs/sPeak MiB
RIFT-S2197.27/198.9940.512979.4
RIFT-S469.99/70.82114.181625.2
SimpleDiff-S2176.61/178.0945.252779.4
Changer control135.62/136.7758.942498.4
ChangeDINO control152.57/160.6551.972529.7

Inputs are 256 × 256 pairs; 50 warmup and 200 timed forwards; TF32 enabled. Latency is per batch; memory is peak allocated; disk I/O/stitching are excluded. Relative to SimpleDiff-S2, RIFT-S2 adds 0.89% parameters, 3.56% counted operations, 11.7% latency and 7.2% memory. Relative to S4, latency/memory are 2.82×/1.83×: denser features incur a substantial cost.

The originally reported S2–S4 comparison was based on the single-seed validation results, whereas the newly added comparison is based on three-seed held-out test evaluations. Therefore, the two sets of differences are not directly interchangeable.

Questions For Authors

Q1. How much improvement comes from fine-tuning?

For the same three seeds with all four method–adaptation cells complete:

QuantityF1 difference (pp)
Fine-tuning gain, RIFT+0.902
Fine-tuning gain, ChangeDINO control+1.119
RIFT advantage with both frozen+0.643
RIFT advantage with both fine-tuned+0.427

Fine-tuning helps ChangeDINO more (+1.119 vs +0.902 pp), yet RIFT retains +0.427 pp with both fine-tuned. Unequal adaptation therefore cannot explain the full advantage. The three-seed fine-tuned estimate is +0.427 pp. These results indicate that full fine-tuning does not account for RIFT's advantage. In fact, ChangeDINO benefits more from fine-tuning, while RIFT retains a consistent advantage under both frozen and fully fine-tuned settings.

AuthorsOpenReview ↗

Official Comment Part 2

Q2. Are the baseline results in Table 1 obtained from our own reimplementations or taken from the original papers? Please provide the exact training/evaluation protocol and checkpoint-selection rule for each benchmark.

Response. Table 1 combines results reported in previous publications and our local reimplementations. We clarify their respective sources and experimental protocols below.

(1) Result provenance.

MethodLEVIRLEVIR+S2LSYSUWHUCDDMSRSCD
FC-Siam-DiPPPMPIR
STANetRPPRPIR
SNUNet-CDPPPMPIR
BITPPRRRIR
ChangeFormerRRRRRIR
ChangerRRRRRRR
RS-MambaPPPRPRR
SPMNetPPPRPRR
ConvFormer-CDRRRRRRR
ChangeDINORRRRRRR

P/I/M: Results transcribed from previously published comparison tables in PhyUnfold-Net, IPACN, and Mask View, respectively. These results were neither retrained nor re-evaluated by us.

R: Results obtained from our local reproductions under a common training protocol, following as closely as possible the corresponding P/I/M settings to ensure fair comparisons.

For P/I/M entries, we directly adopt the results reported in the corresponding publications without retraining these methods. Their training configurations, dataset splits, evaluation procedures, and checkpoint-selection rules therefore correspond to those specified in the original papers. We do not claim that these published results were obtained under our unified training protocol.

For R entries, we train the corresponding methods locally, retaining their original architectures and pretrained initializations and using the official dataset splits. We apply a consistent evaluation implementation to these reproductions. Specifically, the training configuration and checkpoint-selection procedure for each method–dataset pair, including input resolution, optimization settings, learning-rate schedule, augmentation, training budget, validation frequency, selection metric, and tie-breaking rule, are documented in Supplementary. Thus, the distinction between published results and locally reproduced results is explicit and traceable.

(2) Unified protocol for the newly added matched architectural comparisons.

In addition to the heterogeneous benchmark results in Table 1, we conducted controlled comparisons to isolate the architectural contributions of RIFT-CD. For these newly added matched experiments, all compared models use the same DINOv3 ConvNeXt-L/LVD1689M initialization, 256 × 256 paired crops, and an effective batch size of 16.

We train the models using AdamW for 16,000 updates, with 600 warmup updates followed by cosine learning-rate decay and a weight decay of 0.05. The learning rates for the head, stem, and backbone are 2 × 10⁻⁴, 2 × 10⁻⁵, and 6 × 10⁻⁶, respectively. All models use BCE+Dice loss with an auxiliary BCE weight of 0.2 and the same geometric, temporal-swap, photometric, channel-dropout (0.05), and CutMix (0.15) augmentations.

For checkpoint selection, we maintain an EMA model with a decay of 0.999 and evaluate on the validation set every 400 updates. The checkpoint with the highest validation F1 is selected, with the earlier checkpoint preferred in case of a tie. Final evaluation uses the same held-out test scenes and metrics, a fixed threshold of 0.5, and no test-time augmentation or test-set-based model selection.

(3) Scope of the benchmark comparison.

We acknowledge that Table 1 is a literature-level comparison combining published results and local reproductions, rather than a fully protocol-harmonized rebenchmark of all methods. Consequently, it provides a broad comparison with existing approaches, while the newly added matched experiments provide the controlled evidence for the architectural contributions of RIFT-CD. In particular, the matched ChangeDINO and Changer-ConvNeXtL comparisons directly address the effects of backbone adaptation and interaction design, respectively.

Comments on the paper

1 comment

Critical Weaknesses

CW1. Compute cost. S2 improves test localization but costs more latency/memory; Q1/Q3 quantify this. CW2. Training variability. The S2Looking 2 × 2 × 2 factorial now has three seeds (48 models), plus a SYSU test (Q2). CW3. Unseen datasets. Source-only transfer is evaluated without target tuning (Q5).

CW4. Realistic shifts. Registration tests, a seasonal-appearance subset, and natural dataset transfer provide complementary evidence (Q4/Q5).

CW5. Small margins. Q1 reports paired differences, completed seed counts, and the MSRSCD near-tie.

Questions For Authors

Q1. Independent training seeds. RIFT-minus-baseline held-out test differences, paired mean ± sample SD in percentage points (pp), with matched DINOv3 initialization and full fine-tuning:

Dataset / baselineSeed pairsΔF1ΔBand-F1
LEVIR / ChangeDINO3/3+0.427 ± 0.017+1.384 ± 0.063
LEVIR / Changer3/3+0.464 ± 0.104+1.303 ± 0.357
LEVIR / SimpleDiff-S23/3+0.119 ± 0.047+0.197 ± 0.086
LEVIR / RIFT-S43/3+0.390 ± 0.076+0.945 ± 0.064
CDD Model/without_shift / ChangeDINO3/3+0.073 ± 0.004+1.027 ± 0.085
MSRSCD / ChangeDINO3/3−0.012 ± 0.038−0.133 ± 0.380

These held-out estimates are separate from original single-seed validation. Every LEVIR pair improves both metrics. CDD has small positive gains; MSRSCD shows essentially unchanged F1 and lower Band-F1. Seed SD separates retraining variability from scene-bootstrap uncertainty.

Q2. Replication beyond LEVIR. S2Looking test F1 (mean ± SD, three seeds):

EvidenceSelfCross
Raw68.325 ± 0.28468.251 ± 0.271
Normalization only68.403 ± 0.18668.140 ± 0.249
Agreement only68.538 ± 0.12268.231 ± 0.451
Full68.173 ± 0.20068.462 ± 0.233

“Pair-mean replacement only” denotes N=0 and M=1, replacing the pair-mean descriptor with channel-wise agreement without relational L2 normalization. Feedback is independently varied between own-date (Self) and cross-date (Cross), forming a complete 2 × 2 × 2 factorial design.

Only Full has positive mean Cross−Self gains in F1 (+0.288 ± 0.337) and Band-F1 (+0.310 ± 0.237). The Full-versus-Raw F1 interaction is +0.362 ± 0.813 pp; it varies across seeds. Full lowers radiometric/noise message drift at all three levels in every seed on both datasets. The S2Looking test excludes exact single-date images shared with train/validation: 909 scenes, 898 resampling groups; selection uses original validation.

Additional mechanism evidence on SYSU-CD. On 4,000 test pairs using the three seeds, Raw-Cross reduced F1 by 0.1667 pp relative to Raw-Self, whereas Full-Cross improved F1 by 0.0567 pp over Full-Self. This yields a positive interaction effect of +0.2234 pp in F1 and +0.3431 pp in IoU, suggesting cross-branch feedback benefits from full relational evidence.

Q3. Practical cost. Exclusive RTX 4090, 256 × 256 pairs; 50 warmup and 200 timed forwards:

Configurationp50/p95 batch msPairs/sPeak MiB
S2, FP32, batch 132.22/38.8031.931055.8
S2, FP32, batch 8197.27/198.9940.512979.4
S4, FP32, batch 869.99/70.82114.181625.2
S2, BF16, batch 8147.05/148.7954.352792.9

FP32 permits TF32; memory is peak allocated. Timing excludes I/O/stitching. These quantify the practical cost alongside the 3.75× operation count.

Q4. Registration and seasonal appearance.

LEVIR-CD and S2Looking contain imagery acquired across different dates and locations. LEVIR-CD was designed to include seasonal and illumination variations. However, the source imagery may originate from multiple commercial satellites, while the dataset does not provide sufficiently detailed image-level sensor metadata to isolate sensor-specific effects. Therefore, our experiments do not establish cross-sensor robustness.

Registration error in three-seed Full-Cross mean test F1 drops (pp); clean/shifted predictions use the same valid region:

Dataset1 px2 px4 pxRadiometricNoise
LEVIR1.293.869.870.0350.268
S2Looking0.070.321.090.0645.441

RGB gains (1.16, 0.88, 1.04), offset 0.04; noise SD 0.04 on [0,1].

On 31 manually selected LEVIR pairs with visible seasonal appearance changes, RIFT achieves 92.930 F1 versus 92.504 for ChangeDINO and 92.693 for Changer (+0.426/+0.237 pp). This subset analysis is exploratory.

Q5. Source-only transfer. Source-validation checkpoints, threshold 0.5; no target fine-tuning, or threshold tuning. Target F1 is mean over three seeds:

ModelLEVIR→S2LookingLEVIR→WHUS2Looking→LEVIR
Full-Cross40.44%85.56%78.896%
Changer19.64%80.24%35.112%

These real target-image tests show transfer depends on the source–target pair: Full exceeds Changer on all forward targets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.