for the CORL 2020 publication:

Achermann Florian, Kolobov Andrey, Dey Debadeepta, Hinzmann Timo, Chung Jen Jen, Siegwart Roland, Lawrance Nicholas, “MultiPoint: Cross-spectral registration of thermal and optical aerial imagery”, Conference on Robot Learning (CoRL), 2020.

The data was collected using a fixed-wing UAV equipped with two downward facing cameras in ten flights at different times of the day. The RGB camera is a UI- 5261SE Rev. 4 with a 16 mm focal length lens and the thermal camera a FLIR TAU2 19 mm, spectral band 7.5–13.5 μm. In the paper the process of generating the aligned image pairs is described. The dataset is split across flights into a training and test set. An interface class to load the data from the hdf5 files is provided in the multipoint repository.

Each sample in the HDF5 files contains three images (optical, thermal, thermal_raw). All images have a resolution of 640×512, are undistorted and aligned. The optical image is already converted to greyscale and normalized such that 0.0 represents black and 1.0 white. The thermal image is normalized on a per image base such that 0.0 is equal to the 1st percentile and 1.0 to the 99th percentile of all pixel values in that image. The thermal_raw contains the raw pixel values normalized by 16383. The temperature in degrees can be recovered from the pixel values by using the following equation:

temp_deg = pixel_val * 16383 * 0.04 - 273.15

training.hdf5 – 24 GB – Training dataset, contains 9340 image pairs

test.hdf5 – 12 GB – Test dataset, contains 4391 image pairs

labels_training.hdf5 – 10 MB – Keypoint labels for the training dataset.