Aligning Each Object, Not Each Pixel: Change Detection in Unregistered Satellite Pairs

Change detection compares two images of the same place taken at different times. It usually assumes that pixel $(x, y)$ in one image shows the same spot on the ground as pixel $(x, y)$ in the other. In real disaster imagery that assumption fails. Pre- and post-event satellite shots are taken from different angles and altitudes, so even a one-pixel offset at a building edge produces a bright outline of fake "change."

The usual fix is to predict a dense flow field alongside the change map. A dense flow field is a per-pixel displacement map that warps one image onto the other. This fix has a known flaw. E2ECDNet jointly predicts flow and change on unregistered pairs and reaches high recall. Its authors report, however, that its precision falls below cascaded two-stage pipelines. The cause is that the flow network tries to warp a genuinely collapsed building until it resembles the intact one. That distortion creates new false positives, and the authors leave the fix to future work.

This proposal targets that specific failure. The idea is to stop estimating per-pixel motion inside the objects whose change we want to judge. For each building footprint segmented from the pre-event image, the model predicts one rigid displacement. It estimates that displacement from coarse, jitter-tolerant levels of a feature pyramid and from a ring of surrounding context. It then compares fine-resolution features inside the footprint under that single shift.

A single translation per object cannot reshape a destroyed building into an intact one, so the contamination path disappears by construction. Per-footprint translation is also close to the real geometry. Off-nadir viewing shifts elevated rooftops relative to the ground by an amount that depends on the structure, so one shift per building is roughly right while one smooth dense field is not.

Core Claim

In unregistered bi-temporal change detection, most of the precision loss comes from dense warping inside changed objects. Replacing dense within-object flow with one displacement per region of interest (ROI) should recover the precision of cascaded pipelines without losing the recall of joint models. That displacement is estimated from coarse feature-pyramid levels, while comparison happens at fine levels.

The three input directions play different roles. Co-registration sensitivity is the problem being solved. ROI segmentation defines the unit of alignment, not a crop for a classifier. Multi-scale fusion supplies the split of labour: deep, low-resolution levels are stable enough to estimate displacement, and shallow levels are sharp enough to judge change.

The ROI comes only from the pre-event image. This matters because the chicken-and-egg problem stated in the FastPD-MRF work is that alignment needs to know where change is, and change needs alignment. Masking flow with predicted change reproduces that circularity. A pre-event footprint does not, because it is fixed before any change happens.

Relation to Prior Work

Three joint registration-and-change systems motivate this proposal. E2ECDNet refines one dense flow coarse-to-fine through a pyramid and reports the precision loss described above. DiffRegCD predicts dense flow from frozen diffusion-model features, treats displacement prediction as classification over candidate offsets with Gaussian smoothing, and warps features at every level. Dual-PRNet does sequential pyramid warping for brain MRI and notes that interpolation error accumulates across levels. All three warp densely and uniformly across the pyramid, with no object structure. We keep DiffRegCD's classification-over-displacements idea as our estimator, but apply it once per object rather than per pixel.

The closest conceptual precedent is the graph-based joint registration and change detection model. It couples a registration graph and a change graph so that flagged changes relax matching penalties, and it reports Overall Quality rising from 51.3% (sequential) to 74.4% (joint) on Attica imagery. That paper relies on handcrafted features and iterative discrete optimization. We replace the soft relaxation with a hard structural constraint (rigid per-object motion) inside a trainable pyramid, and we condition on a prior that does not depend on the change estimate.

Region-based alignment has also been tried before. SAMReg matches Segment Anything masks between images through pooled prototype vectors and solves for a smooth dense field. It works at one feature scale and targets registration itself, not change. RDDN compares test and template features inside feature-pyramid layers and adds enlarged context boxes around proposals. On fabric it goes from 42.2 to 51.4 mAP versus Faster R-CNN, but it uses rectangles and assumes only small translational gaps. We borrow the context-ring idea from RDDN and the mask-prototype idea from SAMReg, and combine them into a per-footprint displacement estimator.

ROI masking in medicine constrains the design. The mask-weighted pooling CT triage model and DenseSharp+ show that restricting features to a mask removes shortcut cues. Both are single-image methods, and both expose the risk that a bad upstream mask starves the downstream model. Our experiments therefore test deliberately corrupted masks.

Research Program

The model has four parts. It starts with a shared Siamese feature pyramid over both images. A Siamese pyramid is the same backbone with top-down and lateral connections, applied to each image.

The second part is a footprint segmenter run on the pre-event image alone. It is trained on xBD building polygons.

The third part is a per-object displacement head. For footprint $k$, it pools pre-event features over the footprint and a surrounding ring, correlates the result with the post-event feature map over a grid of candidate offsets at the coarsest levels, and refines the offset at the next level:

$$ p_k(d) \propto \exp\Big(\sum_{l \in \mathcal{C}} w_l \,\langle \bar f^{\,pre}_{k,l},\, f^{\,post}_{l}(\cdot + d)\rangle_{M_k}\Big) $$

In this expression, $d$ is a candidate 2D displacement, $\mathcal{C}$ is the set of coarse pyramid levels, $w_l$ are learned level weights, $\bar f^{\,pre}_{k,l}$ is the pre-event prototype of object $k$ at level $l$, and $\langle\cdot,\cdot\rangle_{M_k}$ is correlation restricted to the footprint-plus-ring mask $M_k$. The displacement $d_k$ is the mean of $p_k$, and the spread of $p_k$ serves as an explicit alignment-uncertainty estimate. Uncertainty modelling is a gap DiffRegCD leaves open.

The fourth part is a change and damage head. It compares fine-level features inside footprint $k$ after shifting the post-event features by $d_k$, and it takes the uncertainty as an input. Background pixels still receive a smooth low-resolution dense flow of the E2ECDNet kind, but that flow never governs pixels inside a footprint.

Training uses ground-truth flow and change labels where available, following the synthetic protocol of xBD-E2ECD. We also run two analyses. The first maps where false positives fall, splitting them into the interior of changed objects, the boundaries of unchanged objects, and background. The second measures how fine-level comparison error grows with residual misalignment at each pyramid level, which tests whether the coarse-estimate, fine-compare split is justified.

Baselines

The naive reference is a Siamese pyramid with feature subtraction and no alignment, trained on the same perturbed pairs. It shows how much of the problem is registration at all.

The strongest direct competitor is E2ECDNet, because it is the joint dense-flow model on exactly this benchmark. It reports R@5 of 98.25% on xBD-E2ECD, where R@r is recall that credits detections within r pixels. Its cascaded two-stage baseline reports 96.90%. E2ECDNet's precision is described only as lower than cascaded methods; no reported precision number was found on Lacuna. We will therefore retrain E2ECDNet and the cascade ourselves, with the same backbone, pyramid width, data splits, perturbation distribution, and hyperparameter search budget.

DiffRegCD is the second strong competitor, reporting 72.0 F1 under high perturbation on VL-CMU-CD against 58.9 for ChangeRD. We will run it with diffusion features and also with our backbone, so that feature quality and alignment strategy are not confounded.

For damage classification on registered imagery, PPM-SSNet reports per-class F1 on xBD of 0.90 (no damage), 0.41 (minor), 0.65 (major), and 0.70 (destroyed). We will also retrain it on perturbed pairs.

Two ablations isolate the key ingredient. The first swaps our per-object displacement for dense flow restricted to the same footprints, keeping ROIs, pyramid, and compute identical. The second estimates per-object displacement from fine levels instead of coarse levels.

Beating the baselines means precision at matched recall above both E2ECDNet and the cascade on xBD-E2ECD, R@5 no worse than E2ECDNet's, and a measurable drop in false positives on undamaged-building boundaries.

Why not just…?

Why not just mask the flow loss with the predicted change map? This is the deep analogue of the graph-based joint model and a real threat. Its weakness is circularity. Early in training, predicted change is poor, so the gating either does nothing or suppresses alignment wherever the network is confused. We implement this variant on the same E2ECDNet backbone. If it matches our precision, the object-rigid constraint is unnecessary and we will report that.

Why not just compare features with a spatial tolerance, such as a local max-pool before subtraction, and skip registration? For small jitter this may suffice. Deep-feature comparison is what makes RDDN's TR-Inter and ReDi tolerant to pixel gaps. RDDN, however, notes vulnerability beyond small translations, and xBD-E2ECD uses affine perturbations. We sweep the tolerance window against perturbation magnitude. We expect it to win at small offsets and fail once displacement exceeds the window, but the crossover point is itself a result.

Why not just register offline and then run a standard change network? The file shows cascades losing recall: 96.90% versus 98.25% R@5 for E2ECDNet, and 51.3% versus 74.4% Overall Quality in the graph-based study. Yet the cascade has the best precision. Our claim is precisely that we can match it. If a strong cascade with good per-image registration matches our precision–recall curve, the proposal fails on its central claim.

Proposed Experiments

The main benchmark is xBD-E2ECD, reporting precision, recall, F1, and R@r at several radii across perturbation magnitudes. We add a harder variant that layers random per-footprint offsets, scaled with footprint size, on top of the global affine, to mimic parallax of elevated structures. Damage classification is evaluated on xBD with per-class F1, including the weak minor and major tiers.

For generality, we transfer to VL-CMU-CD, using Segment Anything masks from the first image as pseudo-ROIs in the style of SAMReg. The comparison there is against DiffRegCD's 72.0 F1.

Three checks decide the question. The first is the contamination test. Among dense-flow baselines, flow error inside destroyed buildings should exceed flow error inside undamaged ones. Our per-object displacement error should show no such gap, and closing that gap should explain most of the precision gain.

The second is the ablation of dense flow against rigid displacement inside identical ROIs.

The third is mask robustness. We replace ground-truth footprints with predicted, eroded, dilated, and randomly dropped masks. Precision should degrade gracefully, and dropped footprints should fall back to background dense flow rather than failing silently.

If the gain vanishes under predicted masks, or the change-gated flow variant matches it, the idea is not worth pursuing further.