Research map / Probabilistic and scientific modeling / Dynamical systems

FM-ChangeNet: Learning Change through Pathwise Feature Transport

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent dynamics models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FM-ChangeNet, which treats bi-temporal change detection as feature-space transport, supervising a time-conditioned velocity field over linearly interpolated encoder features with flow, consistency, TV and segmentation losses.

Problem

Change detection from bi-temporal remote sensing images is usually supervised only at the endpoint change mask, leaving the temporal transition between the two observations unconstrained and vulnerable to nuisance variation such as illumination and misalignment.

Approach

Reformulates change detection as learning a time-conditioned velocity field over linearly interpolated intermediate feature states between pre- and post-temporal features (flow matching target f1-f0), and uses the velocity magnitude as a localized change cue alongside segmentation supervision.

Key results

• On LEVIR-CD, FM-ChangeNet reaches F1 92.45 and IoU 85.96, above the listed CDMamba (F1 92.23) and other baselines.
• On WHU-CD, F1 95.44 and IoU 91.28 versus CDMamba F1 95.24 and IoU 90.91 (reported in the table).
• On DSIFN-CD, F1 88.96 and IoU 80.12, above ChangeFormer's F1 86.67 in the reported table.

Surprising findings

• Velocity magnitude alone (normalized) gives competitive segmentation results, e.g., F1 90.09 on LEVIR-CD.
• The paper's own theorem notes the canonical linear-path target f1-f0 is constant in t, so pathwise supervision adds limited mathematical novelty beyond endpoint regression.

Limitations

• Formulation is deterministic and not evaluated under stronger distribution shifts or severe misregistration (stated).
• Efficiency and large-scale constraints are not evaluated (stated).
• Linear interpolation target is constant in t, so the flow target is essentially endpoint displacement.

Experimental setup

Datasets: LEVIR-CD WHU-CD DSIFN-CD

Benchmarks: LEVIR-CD WHU-CD DSIFN-CD

Metrics: Precision Recall F1 IoU OA

Baselines: FC-EF FC-Siam-Diff FC-Siam-Conc IFNet SNUNet SwinUnet BIT ChangeFormer MSCANet Paformer DARNet ACABFNet

Builds on: flow matching ChangeFormer BIT

Closest papers by meaning

• The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.86)
• Towards Generalizable Scene Change Detection (CVPR 2025 · similarity 0.85)
• The STVchrono Dataset: Towards Continuous Change Recognition in Time (CVPR 2024 · similarity 0.84)
• RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-𝑁 Ranking (NeurIPS 2026 · similarity 0.84)
• Multi-View Pose-Agnostic Change Localization with Zero Labels (CVPR 2025 · similarity 0.84)
• Feature Spectrum Learning for Remote Sensing Change Detection (CVPR 2025 · similarity 0.84)