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

Sparse point annotations make complete, reliable change masks difficult to recover. This paper progressively refines SAM2-based pseudo-labels through CNN refinement and teacher–student self-training, achieving the best results among compared weakly supervised approaches on three benchmarks while remaining competitive with several fully supervised baselines.

  • Problem

    Sparse point-level supervision provides limited localization cues, while SAM2 predictions can be unreliable on target change-detection images because of domain gaps and imprecise boundaries.

    Source: Stage I: pseudo-label Generation and… · p. 3

  • Method

    The framework uses SAM2-guided bi-temporal mask selection and CNN refinement to construct pseudo-labels, then applies EMA teacher–student self-training to refresh them progressively.

    Source: Introduction · p. 2, Stage II: Teacher–Student Self-Training · p. 4

  • Results

    On three benchmark datasets, the method achieves the best results among compared weakly supervised approaches and remains competitive with several fully supervised baselines.

    Source: Conclusion · p. 7

  • Takeaways & Limitations

    The reported ablations support combining Stage I pseudo-label refinement with Stage II self-training, which are complementary on the three evaluated datasets.

    Source: Ablation Study · p. 6

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 supervision reduces annotation costs but leaves pseudo-labels sparse and incomplete, making their quality central to point-supervised change detection. The proposed two-stage framework progressively adapts SAM2-derived pseudo-labels and detector predictions through refinement and self-training.

Introduction

  • Sparse change points provide low-cost localization cues but omit object extent, boundary structure, and background information, making pseudo-label quality decisive.The introduction frames point supervision as a weakly supervised alternative to dense pixel-level annotation.Source: Introduction · p. 1
  • Region expansion and CAM propagation can yield incomplete regions, inaccurate boundaries, and false activations, while fixed labels let early errors persist.Foundation-model priors offer object-level segmentation, but single-temporal predictions require adaptation to bi-temporal change detection.Source: Introduction · p. 2
  • SAM2-guided bi-temporal mask selection ranks candidate masks by confidence, change consistency, and point coverage to create change-aware pseudo-labels.The strategy converts generic single-temporal segmentation responses into initial change pseudo-labels.Source: Introduction · p. 2
  • An uncertainty-aware lightweight CNN refinement module improves pseudo-label boundaries and local structural consistency while suppressing unreliable supervision.The refinement addresses imprecise boundaries and local noise in SAM2 priors.Source: Introduction · p. 2
  • An EMA teacher periodically refreshes pseudo-labels from its predictions, enabling a closed loop of model optimization, label updating, and reoptimization.This replaces fixed Stage I supervision with supervision that evolves during training.Source: Introduction · p. 2
  • The framework jointly evolves pseudo-label quality and detector capability through two-stage progressive optimization.Stage I generates and refines pseudo-labels; Stage II uses teacher-student self-training to update supervision during training.Source: Introduction · p. 2

Related Work

Prior weakly supervised change-detection methods depend on pseudo-label quality, while one-shot training and fully supervised foundation-model adaptations leave reliable supervision from sparse points underexplored. The proposed method addresses this gap by progressively updating SAM-generated pseudo-labels through teacher-student self-training.

Weakly Supervised Change Detection in Remote Sensing

  • Weakly supervised change detection still relies on heuristic or coarse pseudo-labels, leaving complete and accurate regions from point supervision underexplored.Prior approaches include CAM localization and SAM refinement in CS-WSCD (Wang, Zhang, and Shi 2023), regional growth in CARGNet, and stronger feature modeling in BARNet and TransWCD.Source: Related Work · p. 2

Pseudo-Label Learning and Self-Training

  • Unlike one-shot pseudo-labeling in remote sensing change detection, the proposed method progressively updates SAM-generated labels through teacher-student self-training.One-shot labels can propagate early errors and limit task adaptation; the method instead uses model feedback to refine them.Source: Related Work · p. 2

Foundation Models for Change Detection in Remote Sensing

  • SAM-CD aligns bi-temporal representations with a convolutional adapter and task-agnostic semantic learning branch; these methods are mainly fully supervised, leaving SAM/SAM2 priors for reliable weak-annotation supervision insufficiently studied.Source: Related Work · p. 3

Methodology

The method progressively adapts SAM2 priors to point-supervised change detection through two stages: generating and refining pseudo-labels, then using teacher–student self-training to refresh them in a closed loop.

Methodology

  • Stage I addresses sparse, noisy point supervision by using SAM2 candidate masks, bi-temporal selection, and CNN refinement to create more reliable change pseudo-labels.The method identifies SAM2’s domain gap and imprecise boundaries as reasons its predictions are unreliable for direct supervision.Source: Methodology · p. 3, Stage I: pseudo-label Generation and… · p. 3

Stage I: pseudo-label Generation and Refinement

  • For each positive point prompt, SAM2 independently generates multiple candidate masks on both temporal images, with each point corresponding to a connected change region.The two timestamps’ candidate sets contain the same number of masks for each prompt.Source: Stage I: pseudo-label Generation and… · p. 3
  • Candidate masks are ranked by SAM2 confidence, change consistency, and point coverage, then fused into component-level soft masks for initial pseudo-labels.Change consistency uses a response map integrating grayscale intensity, Sobel gradients, and high-frequency differences between the images.Source: Stage I: pseudo-label Generation and… · pp. 3–4
  • A siamese encoder-decoder with a shared ResNet-18 backbone refines the soft pseudo-labels, while uncertainty-aware image- and pixel-level weights emphasize reliable regions and suppress ambiguous ones.The uncertainty-aware supervision uses weighted binary cross-entropy and weighted Dice losses; the refinement targets boundary quality and local structural consistency.Source: Methodology · p. 3, Stage I: pseudo-label Generation and… · p. 4

Stage II: Teacher–Student Self-Training

  • Stage II closes the loop: an EMA-updated teacher provides predictions to refresh pseudo-labels, which then supervise the student through the uncertainty-aware loss.The teacher shares the student’s architecture, accumulates task-adaptive knowledge, and pseudo-label updates use a momentum coefficient.Source: Stage II: Teacher–Student Self-Training · p. 4

Experiment

Across WHU-CD, LEVIR-CD, and SYSU-CD, the two-stage method achieves the best overall performance among weakly supervised approaches on most datasets and is competitive with some fully supervised methods. Qualitative and ablation results support the benefits of its pseudo-label refinement and self-training.

Experimental Setup.

  • Experiments use official train-test protocols, 256×256 patches, and centroid-sampled points, comparing 14 methods under their recommended settings.Evaluation uses F1, IoU, precision, recall, and overall accuracy across WHU-CD, LEVIR-CD, and SYSU-CD.Source: Experimental Setup · p. 4, Compare with State-of-the-arts · p. 5

Qualitative analysis.

  • Qualitative comparisons show more complete change regions, clearer boundaries, and fewer false alarms, especially for small structures and complex scenes.The comparisons include fully and weakly supervised methods on benchmark datasets.Source: Figure 3 · p. 5, Qualitative analysis · p. 5

Quantitative analysis.

  • 81.25% F1 and 68.80% IoU on LEVIR-CD; gains over MS-Former are 4.38 and 6.37 points, respectively.Recall is 84.80% versus 84.59% for FC-Siam-Conc and 81.65% for FC-Siam-Diff; average drops versus fully supervised methods are 9.18% in F1 and 12.73% in IoU.Source: Quantitative analysis · p. 5
  • 86.21% F1 and 75.77% IoU on WHU-CD, improving over PGU-CD by 4.39 and 6.53 points, respectively.The method also outperforms fully supervised FC-Siam-Conc, FC-Siam-Diff, and ChangeFormer on this dataset.Source: Quantitative analysis · p. 5
  • 72.28% F1 and 56.59% IoU on SYSU-CD, with gains over MS-Former of 0.11% and 0.14%, respectively.The passage notes that complex scene variations make weakly supervised pseudo-label learning more difficult on SYSU-CD.Source: Quantitative analysis · p. 5
  • The method achieves the best overall performance among weakly supervised approaches on most datasets and is close to or better than some fully supervised methods.The comparison covers three benchmarks and 14 state-of-the-art methods: five weakly supervised and nine fully supervised.Source: Quantitative analysis · p. 5, Table 1 · p. 6

Ablation Study.

  • Ablations support bi-temporal mask selection, uncertainty-aware loss, EMA teacher updates, and refreshing pseudo-labels every 2 epochs.Removing mask selection or uncertainty weighting degrades performance; a non-EMA teacher produces inferior results. The 2-epoch interval achieves 81.25% F1 and 68.80% IoU, while overly frequent or sparse refreshes hurt performance.Source: Table 3 · p. 6, Ablation Study · pp. 6–7
  • Stage I refinement and Stage II self-training each improve over initial SAM2 pseudo-labels, and together yield the best results across all three datasets.Pseudo-label quality also improves progressively through Stage I refinement and Stage II refresh, with higher F1@0.5 and lower MAE.Source: Ablation Study · p. 6, Table 2 · p. 6, Table 4 · p. 7

Conclusion

The conclusion presents a two-stage point-supervised change detection framework that uses foundation-model priors and iterative pseudo-label optimization. It reports the strongest results among compared weakly supervised methods across three benchmarks while remaining competitive with several fully supervised baselines.

  • Stage I uses SAM2 candidate masks, bi-temporal mask selection, and CNN refinement to construct higher-quality soft pseudo-labels; Stage II progressively refreshes them through teacher-student self-training.Source: Conclusion · p. 7
  • The method achieves the best results among compared weakly supervised approaches on three benchmarks while remaining competitive with several fully supervised baselines.Source: Conclusion · p. 7
  • Future work will study improving pseudo-label reliability on challenging scenes and extending the framework to more general forms of weak supervision.Source: Conclusion · p. 7

Related papers

This paper cites

Loading 2609.02171v1…