Optical Navigation Dataset for Solar System Small Bodies

This dataset has been curated for the purpose of training and evaluating a variety of local feature extractors intended for optical navigation in the proximity of Solar System small bodies (SSSBs). It aims to serve as a resource for researchers in the field and it is referenced in the related article titled "CNN-based local features for navigation near an asteroid" [1]. Additionally, the associated Python code for this dataset can be found in [2].

The dataset is a compilation of images obtained from four distinct space missions focused on SSSBs, specifically NEAR Shoemaker (Eros) [3], Hayabusa (Itokawa) [4], Rosetta (67P/Churyumov-Gerasimenko) [5, 6], and OSIRIS-REx (Bennu) [7]. It also incorporates synthetic data generated through the utilization of a Bennu shape model [8] and OpenGL-based rendering software [9, 10]. Access to mission-specific images is available through the NASA Planetary Data System (PDS), and for the Rosetta mission, via the ESA Planetary Science Archive [11].

The prefix rot- has been applied to subsets in which images have been pre-rotated to orient the SSSB's rotation axis upwards within the image frame. These subsets are primarily intended for training purposes and encompass image pairs with pixel correspondences that can be found in the aflow directory. Pixel correspondences are stored as 16-bit PNG images, where the G- and B-channels respectively represent the x and y image coordinates. To facilitate data compression and storage, a fixed scaling coefficient of 8 has been employed to convert the pixel correspondence float array into a 16-bit integer array to be used by the PNG compression. These pixel correspondence files can be loaded using the navex.datasets.tools.load_aflow function from [2].

On the other hand, subsets designated with a -d postfix include depth information (*.d files) and are exclusively employed during the evaluation of the proposed feature extractors. The depth data is stored as scaled grayscale 16-bit integer arrays using PNG compression. A custom additional header accompanies these images, providing two 32-bit float values, namely the subtracted offset v₀ and the scale multiplier s utilized in the calculation of image pixel values as v' = (v - v₀)·s. To access the depth data as a 32-bit float array, researchers can utilize the navex.datasets.tools.load_mono function from [2].

Please note that the file paths in e.g. rot-cg67p-osinac.tar and cg67p-osinac-d.tar archives are the same, so you need to either rename the extracted folder, extract them to different folders, or only extract the archive that you need.

For clarity, it should be noted that subsets lacking the aforementioned pre- or postfixes do not contain paired images and consequently lack pixel correspondences. These subsets were exclusively used for feature extractor training in [1].

The dataset also includes *.ckpt files, which are the trained feature extractor models referred to in [1]. More details about how to use them can be found in [2].