[Paper Review] Guided Anisotropic Diffusion and Iterative Learning for Weakly Supervised Change Detection

Rodrigo Caye Daudt, Bertrand Le Saux+2 more|arXiv (Cornell University)|Apr 17, 2019

Remote-Sensing Image Classification46 references34 citations

TL;DR

The paper introduces an iterative training scheme that cleans noisy parcel-level ground truth for change detection and a guided anisotropic diffusion (GAD) post-processing method to improve semantic segmentation by transferring edge information from input images.

ABSTRACT

Large scale datasets created from user labels or openly available data have become crucial to provide training data for large scale learning algorithms. While these datasets are easier to acquire, the data are frequently noisy and unreliable, which is motivating research on weakly supervised learning techniques. In this paper we propose an iterative learning method that extracts the useful information from a large scale change detection dataset generated from open vector data to train a fully convolutional network which surpasses the performance obtained by naive supervised learning. We also propose the guided anisotropic diffusion algorithm, which improves semantic segmentation results using the input images as guides to perform edge preserving filtering, and is used in conjunction with the iterative training method to improve results.

Motivation & Objective

• Address the challenge of learning accurate pixel-level change detection from noisy parcel-level ground truth data.
• Develop an iterative training framework that alternates between network training and data cleaning to improve model performance.
• Introduce Guided Anisotropic Diffusion (GAD) to fuse guide image edges into segmentation predictions and enhance boundary accuracy.

Proposed method

• Propose an iterative training scheme that alternates network training with predictions-based data cleaning to mitigate non-random label noise (NNAR) in HRSCD.
• Merge network predictions with original reference data using three strategies to handle disagreements and preserve trustworthy training samples.
• Introduce Guided Anisotropic Diffusion (GAD) that applies edge-preserving diffusion guided by input images (and multiple guide images) to refine segmentation predictions.
• Extend anisotropic diffusion to RGB and multiple guides by computing diffusion coefficients from combined gradient information across channels and guides.
• Compare GAD with Dense CRF as post-processing, and show GAD better preserves edges while smoothing incorrect labels.

Experimental results

Research questions

• RQ1 Can an iterative training approach improve change detection when training data contain structured, non-random label noise?
• RQ2 Does incorporating image-guided edge-preserving diffusion improve pixel-level change detection predictions over naive supervision?
• RQ3 How do different strategies for merging noisy ground truth with model predictions affect final performance?
• RQ4 Is Guided Anisotropic Diffusion more effective than Dense CRF in this iterative weakly supervised setting?

Key findings

• An iterative training scheme with data cleaning outperforms naive supervised training on noisy parcel-level ground truth for change detection.
• Referencing original ground truth data at each hyperepoch is essential to avoid performance degradation.
• Guided Anisotropic Diffusion (GAD) outperforms Dense CRF in later training stages for quantitative metrics in this setup.
• GAD effectively transfers edges from guide images into the segmentation predictions, improving boundary accuracy.
• Using GAD in conjunction with the iterative training scheme yields the best results among the methods tested.