H2OPM Image Registration Dataset

<p>The H2OPM Image Registration Dataset is a dataset for the evaluation of (groupwise) registration methods.&nbsp; It consists of 8 reference ortho-photo maps (OPMs) of urban and non-urban areas in Austria. For each reference image,&nbsp; 3-11 historical aerial images captured&nbsp; between May 1943 and May 1945 are available, leading to a total of&nbsp; 42 image pairs. In order to evaluate the registration accuracy,&nbsp; manually selected ground truth correspondences&nbsp; between the historical images and the OPM are provided.</p> <p>Technical Details</p> <p>All images have been scale-normalized to a spatial resolution of 1m . However, for the historical images scale information could only approximated, the actual scale of the images might be actually off up to 30%.</p> <p>Reference OPM images are named &ldquo;ref_i.png&rdquo; (i=reference image number, 1..8).<br> Corresponding historical images are named &ldquo;image_i_j.png&rdquo; (j=historical image number). For example, &ldquo;image_2_4.png&rdquo; is the fourth out of five historical images that could be registered to&nbsp; reference image &ldquo;ref_2.png&rdquo;.</p> <p>Ground truth control points are given in the text files &ldquo;CP_i_j.txt&rdquo;. Each row of the text file&nbsp; describes one control point pair with values x1,y1,x2,y2. The origin of the coordinate system is the&nbsp; top left of the image, x-axis points right, y-axis points down.</p> <p>The dataset is used for evaluation in the following paper:</p> <p><strong>Zambanini S. </strong>&ldquo;<a href="https://doi.org/10.1016/j.patcog.2019.01.024">Feature-based groupwise registration of historical aerial images to present-day ortho-photo maps</a>&ldquo;, <em>Pattern Recognition</em>, 90:66-77, June 2019. (<a href="https://arxiv.org/abs/1811.09081">arXiv</a>)</p> <p>Please cite this paper when publishing results on the dataset.</p>