Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Plug-and-Play DISep: Separating Dense Instances for Scene-to-Pixel Weakly-Supervised Change Detection in High-Resolution Remote Sensing Images

Zhenghui Zhao, Chen Wu, Lixiang Ru, Di Wang, Hongruixuan Chen, Cuiqun Chen

TL;DR

The paper tackles instance lumping in scene-level weakly-supervised change detection when changes are densely distributed. It introduces Dense Instance Separation (DISep), a plug-and-play three-step framework using high-pass CAMs for instance localization, connectivity-based retrieval for instance IDs, and a pixel-to-centroid separation loss to enforce intra-instance coherence, with no inference cost. DISep achieves state-of-the-art results across five remote-sensing datasets and seven baseline WSCD methods, and also improves fully-supervised CD baselines, demonstrating strong practical impact and transferability. The work highlights the value of incorporating instance-level structure under weak supervision to improve pixel-level change quantification in high-resolution remote sensing.

Abstract

Existing Weakly-Supervised Change Detection (WSCD) methods often encounter the problem of "instance lumping" under scene-level supervision, particularly in scenarios with a dense distribution of changed instances (i.e., changed objects). In these scenarios, unchanged pixels between changed instances are also mistakenly identified as changed, causing multiple changes to be mistakenly viewed as one. In practical applications, this issue prevents the accurate quantification of the number of changes. To address this issue, we propose a Dense Instance Separation (DISep) method as a plug-and-play solution, refining pixel features from a unified instance perspective under scene-level supervision. Specifically, our DISep comprises a three-step iterative training process: 1) Instance Localization: We locate instance candidate regions for changed pixels using high-pass class activation maps. 2) Instance Retrieval: We identify and group these changed pixels into different instance IDs through connectivity searching. Then, based on the assigned instance IDs, we extract corresponding pixel-level features on a per-instance basis. 3) Instance Separation: We introduce a separation loss to enforce intra-instance pixel consistency in the embedding space, thereby ensuring separable instance feature representations. The proposed DISep adds only minimal training cost and no inference cost. It can be seamlessly integrated to enhance existing WSCD methods. We achieve state-of-the-art performance by enhancing {three Transformer-based and four ConvNet-based methods} on the LEVIR-CD, WHU-CD, DSIFN-CD, SYSU-CD, and CDD datasets. Additionally, our DISep can be used to improve fully-supervised change detection methods. Code is available at https://github.com/zhenghuizhao/Plug-and-Play-DISep-for-Change-Detection.

Plug-and-Play DISep: Separating Dense Instances for Scene-to-Pixel Weakly-Supervised Change Detection in High-Resolution Remote Sensing Images

TL;DR

The paper tackles instance lumping in scene-level weakly-supervised change detection when changes are densely distributed. It introduces Dense Instance Separation (DISep), a plug-and-play three-step framework using high-pass CAMs for instance localization, connectivity-based retrieval for instance IDs, and a pixel-to-centroid separation loss to enforce intra-instance coherence, with no inference cost. DISep achieves state-of-the-art results across five remote-sensing datasets and seven baseline WSCD methods, and also improves fully-supervised CD baselines, demonstrating strong practical impact and transferability. The work highlights the value of incorporating instance-level structure under weak supervision to improve pixel-level change quantification in high-resolution remote sensing.

Abstract

Existing Weakly-Supervised Change Detection (WSCD) methods often encounter the problem of "instance lumping" under scene-level supervision, particularly in scenarios with a dense distribution of changed instances (i.e., changed objects). In these scenarios, unchanged pixels between changed instances are also mistakenly identified as changed, causing multiple changes to be mistakenly viewed as one. In practical applications, this issue prevents the accurate quantification of the number of changes. To address this issue, we propose a Dense Instance Separation (DISep) method as a plug-and-play solution, refining pixel features from a unified instance perspective under scene-level supervision. Specifically, our DISep comprises a three-step iterative training process: 1) Instance Localization: We locate instance candidate regions for changed pixels using high-pass class activation maps. 2) Instance Retrieval: We identify and group these changed pixels into different instance IDs through connectivity searching. Then, based on the assigned instance IDs, we extract corresponding pixel-level features on a per-instance basis. 3) Instance Separation: We introduce a separation loss to enforce intra-instance pixel consistency in the embedding space, thereby ensuring separable instance feature representations. The proposed DISep adds only minimal training cost and no inference cost. It can be seamlessly integrated to enhance existing WSCD methods. We achieve state-of-the-art performance by enhancing {three Transformer-based and four ConvNet-based methods} on the LEVIR-CD, WHU-CD, DSIFN-CD, SYSU-CD, and CDD datasets. Additionally, our DISep can be used to improve fully-supervised change detection methods. Code is available at https://github.com/zhenghuizhao/Plug-and-Play-DISep-for-Change-Detection.
Paper Structure (30 sections, 15 equations, 9 figures, 8 tables)

This paper contains 30 sections, 15 equations, 9 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Weakly-Supervised Change Detection
  4. Instance Estimation with CAMs
  5. Similarity Metric Optimization
  6. Dense Instance Separation
  7. Preliminary
  8. Instance Localization via High-Pass CAMs
  9. Instance Retrieval by Connectivity Search
  10. Instance Separation with Intra-Instance Consistency
  11. Experiments
  12. Experimental Setup
  13. Datasets
  14. Evaluation Protocol
  15. Baselines
  16. ...and 15 more sections

Figures (9)

  • Figure 1: Motivation for our DISep. We show the change predictions in dense instance scenarios, comparing the baseline method (e.g., TransWCD) against the enhanced method with our DISep. In the baseline method, changed instances tend to merge together, resulting in instance lumping. Our DISep successfully separates these merging instances.
  • Figure 2: Prevalence of dense instance distribution in change detection. We present statistics on the instance distribution within the WHU-CD and LEVIR-CD datasets. (a) WHU-CD: 47.77% of image pairs contain multiple instances. (b) LEVIR-CD: It exhibits a higher density. 79.53% of the examples have multiple instances, and 30.26% contain more than 10 instances. (c) DSIFN-CD: It exhibits 62.62% of the paired image examples involving multiple instances.
  • Figure 3: Overview of our DISep. First, we obtain instance localization masks from the CAM using a high-pass threshold. Then, we implement instance retrieval through connectivity search and acquire instance identity masks. Finally, we employ a separation loss to guide an intra-instance pixel online clustering in feature space, according to the instance identity masks. Note that the high-pass threshold is used for optimization to select reliable samples, not for CAM prediction generation.
  • Figure 4: Three cases of changed instance retrieval. The square with a red border indicates the current pixel. (a) The pixel belongs to a new changed instance. (b) The pixel belongs to an already existing changed instance. (c) The pixel connects two existing changed instances, merging them into one.
  • Figure 5: Examples of the WHU-CD, LEVIR-CD, DSIFN-CD, SYSU-CD, and CDD datasets.
  • ...and 4 more figures