fidoptETH3D high-res multi-view
Two DSLR photographs of one indoor or outdoor scene from nearby viewpoints: the middle photograph of an ETH3D high-res multi-view training scene (fixed) and the one sharing the most keypoints with it (moving).
10 test pairsscenesame patientfixed PHOTO, moving PHOTO2D, distances in pxdelivered alignedresearch only
Read before comparing (8)
- Scenes. Thirteen indoor and outdoor places photographed for the benchmark's training data with a DSLR: a courtyard, a delivery area, building fronts, offices, a meadow, a playground, pipes, a terrace and model terrains, 14 to 76 photographs each. relief and relief_2 show two reliefs in one hall of ETH Zurich's main building, so relief_2 is left out; pipes was photographed between two runs of terrains in the same basement corridor, so it is left out too. The setting reads "same patient" for one place seen twice.
- Pairs. The scene's middle photograph by name and the photograph sharing the most of its keypoints in the provider's COLMAP reconstruction (made from the images alone, no laser scan), a rule set on one scene before the draw: one pair a scene, photographs taken a few steps apart, with parallax.
- Truth. The provider registered the photographs to laser scans of each scene (Faro Focus X 330) and rendered a depth map an image, occlusions and masked regions accounted for. The corpus carries one pixel with a depth in each of 16 x 16 cells of the fixed image through its depth, the two poses and the fisheye camera model into the moving image, kept where the moving image's depth within 3 pixels agrees within 5%: 66 to 250 points a pair (kicker the fewest). Measured by a laser and made into points by programs, no person checking: graded automatic and ranked apart.
- Images. The JPEG photographs as the camera saved them, red, green and blue a pixel, 6048 x 4032, with their fisheye distortion. No mask. Distances are in pixels from 0 at the first pixel's centre; the rTRE divides them by the fixed image's diagonal, 7269 px.
- Alignment. The photographs come unaligned, each corner at the origin, and get no starting transform: viewpoint and depth differ, which no translation or rigid fit from the intensities takes out.
- Baseline. With no registration the 10 test pairs score a TRE of 843 px on average (288 to 1658), rTRE 0.116, terrains closest and playground farthest.
- Overlap. The benchmark's two-view and many-view sets film some of these places with another camera rig; they are not in the corpus. No other dataset holds ETH3D.
- Outside numbers. None compare: the benchmark scores multi-view stereo reconstructions against the laser scans, not points between two photographs.
No method has a result on every pair of this dataset yet.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 10 test scenes drawn with seed 2026 among the 11 independent scenes of the high-res multi-view training data, the other one for development. One pair a scene. The benchmark's test scenes keep their laser scans hidden.
- Truth
- landmarks by a program pixels of the fixed image, one in each of 16 x 16 cells where the laser scan gives a depth, carried into the moving image by the provider's depth map and camera poses and kept where the moving image's depth shows the point is seen
- Source
- www.eth3d.net
www.eth3d.net/datasets
www.eth3d.net/documentation
www.eth3d.net/data/schoeps2017cvpr.pdf - Licence
- CC BY-NC-SA 4.0
1 pair for tuning, never tested
facade