You might also be interested in our newer companion POLAR Traverse Dataset released in 2023!

Description

The Polar Optical Lunar Analog Reconstruction (POLAR) dataset seeks to recreate the imaging conditions at the poles of the Moon for stereo vision evaluation. The dominant factors of appearance at the Lunar poles are oblique sunlight, natural terrain, regolith reflectance, and the absence of atmospheric scattering. Lighting is therefore very harsh, resulting in long cast shadows and high dynamic range (HDR) conditions due to contrasting shadowed and illuminated regions. The surface appearance at the Lunar poles is part of a family of optical environments which are prevalent on airless bodies throughout the solar system, but rarely encountered here on Earth.

In developing stereo vision capability for a proposed NASA mission to the lunar poles, we were surprised to learn that there was little prior work in the area and almost no publicly available rover-relevant data on this topic. Despite the proliferation of robots and the wealth of terrestrial stereo datasets, few in the robotics community have looked at imaging specific to airless planetary bodies. In this work, we sought to develop a common framework for stereo vision researchers: (1) an understanding and reproduction of polar/airless optical conditions in a laboratory setting, (2) a library of stereo images for algorithm development complete with ground truth, and (3) metrics and standards for comparative evaluation.

For experimentation, analog terrains were constructed in a laboratory setting using a regolith simulant testbed and populated with rocks and craters. Each terrain was then illuminated with a constellation of oblique sun sources and imaged with a stereo rig from several viewpoints. Finally, a high resolution LIDAR scanner provided ground truth geometry at sub-2mm accuracy for each image pixel. The complete dataset comprises 12 terrain experiments - featuring over 2500 HDR stereo pairs and 12GB of sensor data, supplemented with ground truth, radiometric calibrations, and registration information.

Changelog:

Attribution

Citation:

Acknowledgments:

Documentation and Support

Primary Dataset

Terrain 1
	Description: This should be an easy hazard detection case. Features a distribution of rocks that are mostly large and prominent, with ample spacing in between. The surface has been prepared with a texture that has undulating features but is relatively smooth. There are 15 rocks in this terrain. Download: 730 MB [.zip]
Terrain 2
	Description: Using the same rocks as Terrain 1, but a giant rock has been introduced into the foreground. The idea is to test a large area of cast shadow and HDR in the form of a prominent, bright foreground object. There are 16 rocks (one giant, and 15 large/medium). Download: 911 MB [.zip]
Terrain 3
	Description: Using the same large rocks as Terrain 1, but small and medium rocks have been added to the scene in a uniform manner. While derived from Surveyor 7 (S7) counts, this distribution is weighted toward larger rocks in order to test hazard performance. There are 28 rocks in this distribution. Download: 917MB [.zip]
Terrain 4
	Description: Using the same distribution as Terrain 3, but rocks have been randomly shuffled to different locations. This distribution is weighted toward larger rocks in order to test hazard performance. One side of the terrain has a greater density than the other. The areal placement of large rocks in this scene makes it a good test of left-right occlusions in stereo. There are 28 rocks in this distribution. Download: 923 MB [.zip]
Terrain 5
	Description: This is an S7 sampled distribution centered on small-medium sized rocks. A cluster of larger rocks appears in the corner where it is easily visible from multiple views. This scene tests hazard detection sensitivity near the 10cm range as well as accurate detection of larger rocks when they are possibly arranged with complex shadows. The surface of the regolith is also contains several small negative features (e.g. pitted, craters) Download: 919 MB [.zip]
Terrain 6
	Description: This is a featureless terrain which is thought to be the hardest case for stereo correlation. The terrain lacks rocks, though the surface has smooth undulations which may provide false positives. The surface has been dusted over, however some brushing and natural texture features remain.   Download: 925 MB [.zip]
Terrain 7
	Description: Large rocks are arranged in a “checkerboard” manner on one side of the sandbox. Small rocks are arranged randomly on the opposite side. The side with the small rocks is meant to simulate a rough-but-traversible area. This scene tests complex shadows between the large rocks and also shadows cast from the large rocks onto the the traversible side. There are 19 rocks in this distribution. Download: 938 MB [.zip]
Terrain 8
	Description: This terrain features a field of small and medium sized rocks centered on a large boulder. The large boulder is actually a cluster of rocks stacked together. This scene is designed to test the effect of a (central rather than peripheral) large cast shadow on mostly passable terrain. As our large rock has complex surface geometry with self-occlusions, it is also a good feature to test stereo reconstruction.  There are 16 rocks in this distribution, including one “giant” rock pile. Download: 927 MB [.zip]
Terrain 9
	Description: This is a distribution of rocks that are at or near the 10cm hazard limit. The rocks are spaced far apart, with none touching. This should be a simple, clear test case for hazard detection. There are 20 rocks in this distribution. Download: 926 MB [.zip]
Terrain 10
	Description: This terrain uses the same rock distribution as Terrain 9 with a different placement scheme. Various rocks are clustered in an area of higher density where many are almost touching, while areas of the terrain have no rocks. The terrain is intended to test the effect of areal density of features on hazard detection/reconstruction accuracy. There are 20 rocks in this distribution. Download: 917 MB [.zip]
Terrain 11
	Description: This scene features a large negative obstacle which is meant represent a smooth "fresh" crater. There are some medium-sized rocks along the rim and periphery. The relief of the crater is exaggerated (particularly the rim) in order to also test detection of steep slopes at the particular angle of view. The shape of the crater produces interesting interior shadows and inter-reflected light. The smooth slopes and terrain provide a correlation challenge case. There are 9 rocks in this distribution and one crater. Download: 903 MB [.zip]
Terrain 12
	Description: The fresh crater in Terrain 11 is eroded to create an “old” crater. The rim is reduced in height and sharpness. Large rocks lie collapsed along one wall and covered in regolith while the other side lacks rocks. The regolith has been surfaced with pockmarks to simulate rough texturing from small rocks and accumulated micro-impacts. There are 21 rocks in this distribution and one crater. Download: 918 MB [.zip]

Dataset Extras

Terrain 13
	Description: This is a small sandbox using the LHT regolith simulant which is prohibitively expensive for use in bulk. In comparison to the JSC-1A simulant used in the other terrains, this material is lighter and closer in appearance to the regolith at the lunar poles. Three small rocks (~10cm) have been placed in the scene, and while the light clocks in azimuth. This data is provided as-is and is not well calibrated. Download: 202 MB [.zip]
Radiometric Calibrations
	Description: Radiometric camera response functions used to convert pixel value to scene irradiance. Calibration images, data, and curves in lookup table format. Download: 16 MB [.zip]
Organized GroundTruth Point Clouds
	Description: Organized point clouds are 3-channel images mapping 3D LIDAR scan points to projected locations in a rectified coordinate system. Useful to compare pixel-wise errors with intensity images. Download:380 MB [.zip]
Dataset Tools and Code
	Description: Matlab functions and scripts for programmatically parsing, loading and manipulating the data. Note: Due to software release regulations, we are unable to provide these files at this time.