PaperlayerView exact revision on arXiv

Source-linked AI summary

Progressive Pseudo-Label Optimization for Point-Supervised Change Detection

Hailong Ning, Hao Wang, Yimeng Wang, Tao Lei, Renwei Dian, Asoke K. Nandi

arXiv:2609.02171v1cs.CV

TL;DR

Point supervision reduces annotation costs for change detection but often yields incomplete and noisy pseudo-labels. The proposed two-stage framework uses SAM2, mask selection, CNN refinement, and EMA-based self-training to progressively adapt supervision. Across three benchmark datasets, it outperforms previous weakly supervised approaches on most benchmarks and remains competitive with several fully supervised methods.

  • Problem

    Point annotations reduce labeling costs but have limited spatial coverage, producing incomplete and noisy pseudo-labels for dense change detection.

  • Method

    A two-stage framework generates SAM2 candidate masks, selects and refines change pseudo-labels, then uses EMA-based teacher-student self-training with periodic refreshes.

  • Results

    The method achieves the best results among compared weakly supervised approaches on three benchmark datasets while remaining competitive with several fully supervised baselines.

  • Takeaways & Limitations

    Progressive pseudo-label optimization jointly improves supervision quality and task-specific detector adaptation during training.

Abstract

from arXiv · show

Point-supervised change detection (PS-CD) aims to identify pixel-level changes between bi-temporal images using only sparsely annotated points. Although point annotations substantially reduce labeling costs, their limited spatial coverage often results in incomplete and noisy pseudo-labels. To address this issue, we propose a two-stage framework that introduces SAM2 priors into PS-CD and progressively adapts them to the target task. In Stage I, SAM2 generates object-aware candidate masks from point annotations on the bi-temporal images, and a bi-temporal mask selection strategy is designed to convert generic segmentation responses into more reliable change pseudo-labels. Subsequently, a lightweight CNN refinement module with an uncertainty-aware loss is employed to improve boundary quality and local structural consistency. In Stage II, we construct a teacher-student self-training framework in which the teacher is updated by exponential moving average and periodically refreshes the pseudo-labels. This design establishes a closed-loop optimization process that alternates between pseudo-label refinement and model re-optimization. Experiments on three benchmark datasets, including WHU-CD, LEVIR-CD, and SYSU-CD, demonstrate that the proposed method outperforms previous weakly supervised approaches on most benchmarks and remains competitive with several fully supervised methods.

Introduction

Point-supervised change detection reduces annotation costs but provides insufficient spatial information for dense prediction, leaving pseudo-labels incomplete and noisy. The proposed framework progressively refines SAM2-derived pseudo-labels and adapts them through iterative self-training.

  • Point annotations lower labeling costs but provide little information about object extent, boundaries, and background regions.
  • Existing region-expansion and CAM-propagation methods can produce incomplete regions, inaccurate boundaries, and false activations.
  • SAM2-guided bi-temporal mask selection combines SAM confidence, change consistency, and point coverage to create change-aware soft pseudo-labels.
  • An uncertainty-aware CNN refinement module improves pseudo-label boundary quality and local structural consistency while suppressing unreliable supervision.
  • The framework replaces fixed Stage I supervision with EMA-based teacher updates and periodic pseudo-label refreshes during training.
  • Progressive pseudo-label optimization jointly evolves pseudo-label quality and detector capability through iterative optimization.

Related Work

Weakly supervised remote-sensing change detection reduces the burden of dense annotation through image-, point-, and region-level supervision. Foundation-model methods offer strong segmentation priors, but domain gaps and the bi-temporal setting require task-specific adaptation.

  • Dense pixel-level annotations are costly and labor-intensive, motivating weakly supervised image-, point-, and coarse-region settings.
  • Point-level supervision is attractive because it provides explicit localization cues with substantially less annotation effort than dense change masks.
  • Weakly supervised change detection commonly converts limited annotations into pseudo-labels before training a dense predictor.
  • CS-WSCD uses CAMs to coarsely localize changed regions and SAM to refine ambiguous areas.
  • SAM provides strong zero-shot segmentation, but direct remote-sensing application is limited by the domain gap and the bi-temporal nature of change detection.

Methodology

The method builds initial change pseudo-labels from SAM2’s bi-temporal candidate masks, refines them with uncertainty-aware CNN supervision, and progressively updates them through teacher-student self-training. This closed-loop design alternates pseudo-label refinement with task-specific detector optimization.

  • Stage I: Pseudo-label Generation: SAM2 generates multiple candidate masks independently on each temporal image from sparse point prompts corresponding to connected change regions.The candidate sets contain N masks for each prompt point at both timestamps.
  • Stage I: Pseudo-label Generation: Bi-Temporal Mask Selection ranks candidates using SAM2 confidence, change consistency, and point coverage before fusing selected masks into soft component-level pseudo-labels.Change consistency uses a response map integrating grayscale intensity, Sobel gradients, and high-frequency differences between the images.
  • Stage I: Pseudo-label Refinement: A siamese CNN change detector with a shared ResNet-18 backbone decodes absolute multi-scale feature differences to produce dense change probability maps for pseudo-label refinement.The refinement module targets imprecise boundaries and local texture details in SAM2 outputs.
  • Stage I: Pseudo-label Refinement: Uncertainty-aware supervision assigns image-level and pixel-level confidence weights so reliable pseudo-label regions contribute more while ambiguous regions are suppressed.The loss combines weighted binary cross-entropy and weighted Dice loss under the uncertainty-aware weight w(x).
  • Stage II: Teacher–Student Self-Training: Stage II uses a teacher-student self-training framework in which the teacher is updated as the student’s exponential moving average and generates guidance for pseudo-label refinement.The framework is intended to establish a closed-loop process in which SAM2-initialized pseudo-labels interact continuously with the task-specific change detector.
  • Stage II: Teacher–Student Self-Training: Refreshed pseudo-labels are reused to supervise the student with the uncertainty-aware loss, completing the iterative optimization cycle.The pseudo-label update uses a momentum coefficient to combine labels before and after each refresh step.

Experiment

Experiments on three benchmarks show that the method performs strongly against weakly supervised and several fully supervised approaches. Ablations further indicate that progressive pseudo-label refinement, uncertainty-aware supervision, and EMA self-training improve both pseudo-label quality and detector performance.

  • Benchmark comparison: The method is compared with 14 state-of-the-art methods across WHU-CD, LEVIR-CD, and SYSU-CD under officially recommended settings.The comparison includes 5 weakly supervised and 9 fully supervised methods.
  • Benchmark comparison: 81.25% F1 and 68.80% IoU are achieved on LEVIR-CD, with gains of 4.38 and 6.37 points over MS-Former.The method also surpasses FC-Siam-Conc and FC-Siam-Diff in Recall, reaching 84.80%.
  • Benchmark comparison: 86.21% F1 and 75.77% IoU are achieved on WHU-CD, exceeding PGU-CD by 4.39 and 6.53 points.The method also outperforms several fully supervised baselines in this challenging setting.
  • Benchmark comparison: 72.28% F1 and 56.59% IoU are achieved on SYSU-CD, improving over MS-Former by 0.11% and 0.14%.The reported margin remains positive despite more complex scene variations.
  • Qualitative comparison: Qualitative comparisons show more complete change regions, clearer boundaries, and fewer false alarms, especially for small structures and complex scenes.These observations are reported as consistent with the quantitative results.
  • Ablation study: CNN refinement and Stage II self-training each improve over initial SAM2 pseudo-labels, while the full model performs best across all three datasets.The two stages are described as complementary.
  • Ablation study: Pseudo-label quality progressively improves from initial SAM2 outputs through Stage I refinement and Stage II refresh, with higher F1@0.5 and lower MAE.The analysis attributes this improvement to the supervision signal itself, not only to the final detector.
  • Ablation study: Removing mask selection, uncertainty-aware supervision, or EMA updating causes performance declines, supporting their roles in reliable and stable optimization.EMA updating is linked to more stable supervision and task-specific adaptation.

Conclusion

The paper presents a two-stage point-supervised change detection framework that combines foundation-model priors with iterative pseudo-label optimization. Experiments show strong performance against weakly supervised approaches while remaining competitive with fully supervised baselines.

  • Conclusion: The proposed framework exploits foundation-model priors and iterative pseudo-label optimization for point-supervised change detection.It is organized as a two-stage approach.
  • Conclusion: SAM2 candidate masks, bi-temporal mask selection, and CNN refinement construct higher-quality soft pseudo-labels.These components operate in Stage I.
  • Conclusion: A teacher-student self-training framework progressively refreshes pseudo-labels and improves task-specific adaptation.This constitutes Stage II of the framework.
  • Conclusion: The method achieves the best results among compared weakly supervised approaches while remaining competitive with several fully supervised baselines.The reported experiments cover three benchmark datasets.
  • Conclusion: Future work will study pseudo-label reliability on challenging scenes and extend the framework to more general forms of weak supervision.This defines the stated scope boundary for further investigation.
Loading 2609.02171v1…