Multimodal Image Registration Benchmark: Aerial, Cytological, and Histological Data

The dataset is an evaluation set for multimodal image registration, containing 864 image pairs created from three publicly available 2D datasets. It includes Aerial (Near-Infrared and RGB), Cytological (Fluorescence and Quantitative Phase), and Histological (Second Harmonic Generation and Bright-Field) image pairs, each subjected to four levels of rigid transformations. The data is structured for 3-fold cross-validation, with specific sample counts per fold for each sub-dataset.

Use Cases

• Benchmarking rigid image registration algorithms using the four defined levels of transformation.
• Developing cross-modal feature matching techniques between Near-Infrared and RGB aerial imagery.
• Training models for aligning Fluorescence and Quantitative Phase Images in cytological analysis.
• Evaluating registration methods on Second Harmonic Generation and Bright-Field histological tissue microarrays.
• Studying the generalization of registration models across aerial, cytological, and histological domains.

Strengths

• Provides a structured benchmark with precisely defined transformation levels (4 rigid transformations) for controlled evaluation.
• Integrates three distinct, publicly available datasets covering aerial, cytological, and histological modalities, enabling cross-domain testing.
• The evaluation set contains a specific total of 864 image pairs, and the 3-fold cross-validation structure provides 72 test samples per fold for aerial data and 420 per fold for cytological data.

Limitations

• Key metadata such as total row count, individual dataset sizes, and last update dates are not provided across all sources.
• The dataset is described solely as an evaluation/benchmark set, suggesting it may not be suitable for training large models due to its limited, synthesized nature.
• Column names and specific file formats for the image pairs are unknown, which may complicate initial data handling.

Provenance

Source

Compiled from three publicly available datasets (Aerial, Cytological, Histological) by Jiahao Lu, associated with Uppsala University.

Collection Method

Images were processed by subdividing into patches (e.g., 300x300 px, 834x834 px) and applying synthetic rigid transformations to create an evaluation set.