Beyond Attention: Convolutional Global Context for Remote Sensing Change Detection¶

Conference: ECCV 2026
Paper: Official page · Paper PDF
Code: https://github.com/NUST-Machine-Intelligence-Laboratory/ChangeGCC
Area: Remote Sensing
Keywords: Change detection, globally conditioned convolution, dynamic kernels, cross-scale aggregation, dual-temporal fusion

TL;DR¶

ChangeGCC conditions convolution kernels on scene-level descriptors, then combines intra-frame cross-scale aggregation with symmetric cross-temporal gating to detect remote sensing changes without a global pairwise attention matrix; its Base variant achieves 85.28% IoU on LEVIR-CD with 13.60M parameters.

Background & Motivation¶

Remote sensing change detection compares two observations of the same location and predicts a pixel-level change mask rather than an image-level difference label. Appearance differences are not necessarily physical changes: illumination, seasonality, and radiometric variation can affect large areas, while a newly constructed building may occupy only a few pixels. Convolutional approaches such as FC-Siam and SNUNet preserve spatial structure efficiently, but local evidence can be ambiguous under these conditions. BIT and ChangeFormer introduce long-range interactions to contextualize local decisions, although standard global self-attention incurs quadratic cost in the number of spatial positions.

Linear attention and state-space models reduce asymptotic complexity, but this does not guarantee an inexpensive end-to-end detector. The authors highlight ChangeMamba's additional spatiotemporal relationship module as an example of overhead outside the core sequence operator. Their question is therefore not whether every attention architecture is inefficient, but whether native two-dimensional convolutions can acquire useful global conditioning without explicit all-pairs interactions. A second issue remains even with globally informed features: inconsistent semantics across scales and timestamps can still contaminate the subsequent comparison.

ChangeGCC separates global conditioning within each image from mutual filtering between images. It first uses scene statistics to select convolutional behavior, then consolidates multi-scale semantics before exchanging channel and spatial gates across timestamps. Core idea: let global content determine local filtering, and use cross-scale aggregation followed by symmetric temporal conditioning to distinguish structural changes from appearance variation without global pairwise attention.

Method¶

Overall Architecture¶

The input is a pair of remote sensing images; the output is a dense change-probability map at the original spatial resolution. Each timestamp passes through a four-stage encoder built around Globally Conditioned Convolution, producing a four-level feature pyramid. Dual-Temporal Fusion (DTF) then applies Intra-frame Cross-Scale Aggregation separately to each image, followed by Symmetric Cross-Temporal Interaction at each scale. A lightweight U-Net-like convolutional decoder combines the fused pyramid, restores spatial resolution, and applies sigmoid to produce change probabilities. The decoder is conventional output machinery rather than a separate claimed contribution.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Bi-temporal images"] --> B["Globally Conditioned Convolution<br/>Four feature levels per image"]
    B --> C["Intra-frame Cross-Scale Aggregation<br/>Process each timestamp separately"]
    C --> D["Symmetric Cross-Temporal Interaction<br/>Cross-channel and cross-spatial gates"]
    D --> E["Convolutional decoder and sigmoid"]
    E --> F["Pixel-level change probabilities"]

Training uses paired images and change annotations for end-to-end optimization; inference requires only the image pair. There is no additional text input, teacher network, or test-time optimization loop. “Fully convolutional” does not mean the model lacks pooling or gating: it means that explicit token-to-token attention is not its main global-context mechanism.

Key Designs¶

1. Globally Conditioned Convolution: use scene statistics to select local filtering behavior

A standard convolution uses the same learned spatial kernel regardless of whether the current scene contains dense buildings or open terrain. Globally Conditioned Convolution (GCC) first projects the input features and applies global average pooling to obtain a compact scene descriptor. Lightweight mappings convert this descriptor into sigmoid routing coefficients, which combine a set of learned kernel bases. The central relationship stated in Section 3.1 is:

\[ K(g)=\sum_{i=1}^{M}\alpha_i(g)K_i,\qquad \alpha_i(g)=\sigma(\phi_i(g)). \]

Here, \(g\) is the global descriptor, \(K_i\) are learned kernel bases, \(M\) is their number, and the coefficients depend on the current input. The coefficients use sigmoid rather than a softmax across bases, so they should not be interpreted as probabilities summing to one. The assembled kernel aggregates a separately projected feature representation within a local neighborhood at each position. Global information changes the filtering rule shared across the feature map; the operation does not directly connect every pixel to every other pixel.

This distinction explains both the efficiency and the representational constraint. For fixed kernel support, channel widths, and number of bases, pooling and local filtering scale linearly with spatial size and require no quadratic attention matrix. However, the global descriptor compresses the scene: it does not preserve an independent relationship for every pair of locations. The useful hypothesis is that scene-dependent filtering provides enough context to reduce appearance-driven confusion, not that one GCC layer reproduces arbitrary global attention. The four-stage encoder retains two-dimensional feature layouts throughout and repeats this processing at multiple resolutions for dense localization. The cached main text does not specify the number of kernel bases, exact kernel support, or every block-level setting, so these implementation details should not be guessed.

2. Intra-frame Cross-Scale Aggregation: reconcile semantics before comparing timestamps

High-resolution features retain fine boundaries, while coarser features provide stronger object and region semantics. Comparing timestamps independently at each scale can mix actual temporal differences with insufficient semantic support at a particular resolution. DTF therefore first upsamples all four feature levels of one timestamp to the highest feature resolution using bilinear interpolation. It concatenates them along channels and applies GCC, producing a representation informed by all pyramid levels. The highest feature resolution means the finest level of the feature pyramid, not necessarily the original image resolution. Scale-specific downsampling mappings then project this fused representation back to the four resolutions used by the temporal interaction stage.

Unlike a conventional FPN's progressive top-down pathway, all levels first meet in a common spatial representation, where content-conditioned convolution performs aggregation. The two timestamps remain separate during this step; temporal mixing happens only afterward. The design gives each scale access to complementary semantic and spatial evidence instead of leaving all reconciliation to the final decoder. Its low parameter count should not be confused with negligible computation, because feature fusion takes place at the finest pyramid resolution.

3. Symmetric Cross-Temporal Interaction: let each timestamp filter the other

After scale refinement, each level contains a pair of corresponding feature maps, denoted \(X_1\) and \(X_2\). The module independently computes global average and maximum pooled statistics for each map, transforms them through lightweight bottleneck mappings, adds the responses, and applies sigmoid to obtain channel gates. Average and maximum pooling provide complementary summaries of overall and strongly activated content. Crucially, the gates are exchanged: the channel gate generated from the second timestamp multiplies the first timestamp's features, and vice versa. This is mutual conditioning, not merely a separate self-enhancement block applied to each image.

Spatial conditioning follows channel conditioning. For each channel-modulated map, the module computes channel-wise mean and maximum maps, concatenates them, and predicts a spatial gate with a small convolutional function. These spatial gates are also exchanged between timestamps, after which the two conditioned outputs are added to form the fused representation at that scale. The sequence is therefore cross-channel gating, cross-spatial gating, and symmetric aggregation—not an absolute difference followed by attention. “Symmetric” describes the bidirectional data flow; the main text does not establish whether directional parameters are shared, so strict invariance to timestamp exchange should not be assumed. The supervised objective must teach these gates to preserve useful change evidence: gating alone is not a guaranteed rule for removing pseudo changes. Nor is it a geometric registration mechanism; element-wise modulation does not correct spatial misalignment between images.

A Worked Example¶

Consider two aligned urban images in which lighting changes the appearance of many roofs and one building has genuinely been added. This is an illustrative walkthrough, not an additional quantitative example reported by the paper. Each encoder uses its own scene descriptor to adapt filtering and constructs four levels of features rather than directly subtracting raw pixel intensities. Intra-frame Cross-Scale Aggregation supplements local boundaries with coarser building semantics and redistributes that information to each scale. The timestamps then exchange channel and spatial gates, learning to suppress unhelpful appearance responses while retaining evidence useful for structural change localization. The decoder combines the resulting features into a probability map on the original image grid. Whether the new building is correctly retained depends on learning; the architecture alone does not imply a guaranteed correct mask.

Loss & Training¶

The objective combines pixel-wise binary cross-entropy with a region-overlap Dice loss:

\[ \mathcal{L}=\mathcal{L}_{\mathrm{BCE}}+\lambda\mathcal{L}_{\mathrm{Dice}},\qquad \lambda=1. \]

BCE provides local supervision, while Dice encourages agreement over changed regions rather than treating every pixel decision in isolation. The cached extraction corrupts parts of the expanded BCE and Dice equations, so this note retains only the unambiguous combined objective instead of reproducing damaged expressions. Section 4.1 follows Open-CD preprocessing and uses \(256\times256\) patches for training and inference. Training uses four RTX 3090 GPUs, batch size 8 per GPU, Adam, and 200 epochs; the initial learning rate is \(10^{-5}\) and weight decay is \(10^{-4}\). The cosine schedule includes 20 warmup epochs; the paper also reports a warmup multiplier of 10 and gradient-clipping threshold of 0.5. It calls an Adam setting “momentum 0.999” without giving a complete beta configuration, which should be checked against the implementation rather than treated as a full optimizer specification. Augmentation first applies shared geometric transformations to preserve alignment, then independent photometric perturbations to simulate appearance-driven pseudo changes. T/S/B use encoder depths [2,2,4,2], [2,2,7,2], and [2,3,10,3], respectively; channel widths are [32,64,128,240] for T and [40,80,160,320] for S/B.

Key Experimental Results¶

Main Results¶

The following selection comes from Table 1. IoU and F1 are percentages, with higher values better; lower parameter counts and FLOPs indicate lower resource requirements. These are the paper's dataset evaluation settings: the main text specifies Open-CD patch preprocessing but does not enumerate the train/validation/test split counts, so exact split sizes cannot be supplied from this source. The authors state that baselines use official implementations or reproductions following their configurations under common experimental settings. The costs below are interpreted under the reported \(256\times256\) input protocol, not the \(640\times640\) cost illustration in Figure 1.

Method	Parameters M	FLOPs G	LEVIR-CD IoU ↑	LEVIR-CD F1 ↑	WHU-CD IoU ↑
ChangeMamba	84.70	44.86	84.27	91.37	88.02
ChangeTitans	27.15	30.39	84.36	91.52	88.56
S2ENet	24.92	17.59	84.61	91.67	89.33
ChangeGCC-T	6.34	13.60	84.85	91.80	89.52
ChangeGCC-B	13.60	26.49	85.28	92.06	89.95

Relative to ChangeMamba, Base gains 1.01 IoU percentage points on LEVIR-CD while reducing parameters from 84.70M to 13.60M, directly calculated from Table 1. It is not uniformly cheaper than every comparator: S2ENet's 17.59G FLOPs are below Base's 26.49G, illustrating an accuracy–cost trade-off rather than universal dominance. Tables 2 and 3 report Base IoUs of 75.92% on LEVIR-CD+ and 71.54% on SYSU-CD, versus 75.84% and 71.18% for ChangeTitans. On SAR-CD, Table 4 reports 97.21% IoU for Base and 97.09% for MFPNet; this benchmark contains synthetic changes and is not a real cross-sensor transfer evaluation.

Ablation Study¶

The following subset of Table 5 uses ChangeGCC-B on LEVIR-CD, changing only the indicated component while keeping the other settings fixed. IoU is a percentage and higher is better; these are component replacements within the architecture, not repeated results from complete external models.

Configuration change	IoU ↑	Change from full model, percentage points
Full model	85.28	0.00
Replace GCC with fixed \(3\times3\) convolution	81.80	-3.48
Replace GCC with fixed \(7\times7\) convolution	83.24	-2.04
Remove Intra-frame Cross-Scale Aggregation	84.53	-0.75
Replace cross-scale aggregation with FPN	84.96	-0.32
Replace temporal interaction with SiamDiff	82.06	-3.22
Replace temporal interaction with STRM	84.72	-0.56
Remove Dice term, \(\lambda=0\)	84.97	-0.31

Key Findings¶

Fixed \(7\times7\) convolution still trails GCC by 2.04 IoU points, supporting a benefit beyond simply enlarging the neighborhood. Table 5 does not provide the replacement variants' parameters and FLOPs, so this is not established as an exactly matched-budget comparison.
Cross-scale aggregation adds 0.75 IoU points. Table 6 assigns it 0.19M parameters and 6.09G FLOPs in Base—1.40% and 22.99% of the respective totals—making it parameter-light but computationally substantial.
Replacing temporal interaction with SiamDiff loses 3.22 points. This supports learned mutual modulation in this backbone, not a universal claim that differencing-based detectors are inadequate.
Figure 6 measures throughput on an RTX 3090 with batch size 1, AMP disabled, five warmup runs, and an average over the next ten runs. The cached plot extraction does not preserve reliable curve values, so no exact FPS or speedup factor is quoted.

Highlights & Insights¶

Global information can control an operator rather than directly transport every distant feature. GCC compresses scene context into kernel weights, obtaining adaptive convolution while accepting a loss of spatial information in the descriptor.
Resolve scale inconsistency before temporal disagreement. DTF assigns these problems to successive stages instead of asking the final decoder to handle every type of fusion.
Count operations where they occur. Cross-scale aggregation has few parameters but runs at high feature resolution, so model-file size alone is a poor deployment proxy.

Limitations & Future Work¶

Author-stated limitations: learning with limited supervision and adapting to varied sensing conditions remain open; proposed directions include weak supervision, broader cross-modal settings, and real-time deployment.
Reader interpretation: pooled global descriptors discard spatial arrangement and do not implement independent interactions for arbitrary location pairs. Tasks requiring precise long-range correspondence need separate evaluation.
Scope of generalization: separate experiments on five datasets do not demonstrate zero-shot cross-dataset transfer. Synthetic SAR changes also do not establish real optical-to-SAR transfer.
Deployment evidence: the paper analyzes desktop-GPU throughput and memory, not measured onboard energy or end-to-end latency on UAV or satellite hardware. Resource-constrained deployment remains a potential application rather than a completed validation.
Reproduction and source issues: split details and kernel-base settings are incomplete, and several extracted equations are damaged. The reference attached to LEVIR-CD+ names S2Looking, requiring dataset/citation verification. Some Table 5 IoU/F1 pairs also raise consistency concerns; the ablation selection therefore reports clearly recoverable IoU values without inventing corrected F1 values.
Label-noise evidence: Figure 5 offers qualitative examples of missing annotations, not a controlled noise-rate study with repeated trials.

vs FC-Siam: early Siamese approaches fuse timestamps through differences or concatenation. ChangeGCC first reconciles scales, then conditions each timestamp with gates produced by the other.
vs BIT / ChangeFormer: token interactions supply long-range context, whereas GCC changes local convolution through a global descriptor. Avoiding pairwise costs should not be confused with identical representational capacity.
vs ChangeMamba / ChangeTitans: state-space and neural-memory approaches provide alternative linear-complexity routes. ChangeGCC emphasizes native two-dimensional operations and whole-network cost; practical comparisons still require matching input sizes, precision, and hardware.
vs dynamic convolution: the authors explicitly build on content-adaptive kernel generation. The contribution is better understood as organizing global conditioning and dual-temporal fusion into an effective detector than as inventing dynamic filtering from scratch.

Rating¶

Novelty: 3.5/5 — A coherent combination of global conditioning and symmetric temporal fusion, grounded in existing dynamic-kernel and gating ideas.
Experimental Thoroughness: 4/5 — Five datasets, component ablations, and efficiency analysis; repeated-run statistics and actual edge-device validation remain absent.
Writing Quality: 3/5 — The main mechanism is clear, but dataset citations, some metric pairs, and implementation details need checking.
Value: 4/5 — Useful for efficient remote sensing change detection, especially for evaluating global-context mechanisms together with fusion overhead.