NA-CD: Non-Aligned Change Detection Benchmark with Affine Transformations

266.9 KB of JPG images form the NA-CD dataset, constructed by Zhe Xing to address a lack of public benchmarks for non-aligned change detection. The dataset was created by applying artificial affine transformations—spanning translations, rotations, and scalings—to imagery from source datasets like DOTA 1.5, Google, LEVIR-CD, and WHU-BCD. It was last updated on figshare in April 2026.

Use Cases

• Benchmarking model robustness to geometric distortions based on the introduced affine transformations.
• Evaluating change detection algorithms on non-aligned imagery based on the simulated platform motion and multi-source acquisition.
• Training models to recognize synthetic change types based on the noise, blur, and gray-pixel filling strategies described.
• Testing generalization across different data generation methods based on the four distinct source datasets used.

Strengths

• Explicitly designed to fill a gap in publicly available datasets for non-aligned change detection.
• Simulates a wide range of real-world geometric misalignments through controlled affine transformations.
• Built from multiple established source datasets (DOTA 1.5, Google, LEVIR-CD, WHU-BCD) for generality.
• Released under a permissive CC-BY-4.0 license.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The 266.9 KB size suggests a limited scope, likely a small benchmark set rather than a large training corpus.

Provenance

Source

Constructed by author Zhe Xing from multiple source datasets.

Collection Method

Synthetic generation via affine transformations and region corruption applied to selected source imagery.

Time Range

null

Freshness

Last updated 2026-04-10 05:49:23; freshness should be verified.

Geography

null