fidopt Interactive view
Speckle images (2D), known sub-pixel displacement, ranked apart

2D DIC Challenge 1.0

A speckle pattern's reference image (fixed) and an image of it moved by a known sub-pixel to few-pixel displacement (moving), from the 2D Digital Image Correlation Challenge 1.0.

9 test pairsspecklesame patientfixed PHOTO, moving PHOTO2D, distances in pxdelivered alignedresearch only
Look at the images
Read before comparing (7)
  • Images. Speckle patterns on a flat surface, real or rendered, as digital image correlation measures them: grey, 8 bits, 325 x 487 to 1000 x 2000 px, noise of 0.3 to 8 grey levels, some at low contrast (Samples 2, 4, 7) or with a contrast that changes between the two images (Samples 1, 5). As delivered, spacing 1; distances in pixels from 0 at the first pixel's centre. No mask.
  • Cases. One a speckle pattern, since images of one pattern are not independent: Samples 3 and 3b share one, 5, 9 and 11b another (peaks of cross-correlation 1.0, every other pair under 0.25). Each case pairs the reference with the image the challenge moved most (shifts) or analysed (14, 15). Left out: rotations 8 and 9 (no centre stated), the strain gradients 10, 11 and 11b (their functions' parameters are not published), and the experiments 12, 13, 16 and 17 (no exact truth).
  • Truth. Samples 1 to 7: a uniform shift of 1.0 px in x and y, exact by construction (TexGen, Fourier shift, or a 16-megapixel image binned 10 by 10 from offset rows and columns); TexGen counts y upwards, so Samples 1 and 2 move by -1.0 px in image rows. Sample 14: u(x) = a sine of 0.1 px whose period shrinks along x. Sample 15: v(y) up to 9.6 px with a strain concentration and a step. 14 and 15 come from the challenge's commanded fields, moved to the reference's positions with its own code. A displacement imposed by the image generator, made by no person: graded synthetic and ranked apart.
  • Checked. On every case the moving image sampled through the truth matches the reference better than with no displacement or the reversed one: residual 0.43 against 33.6 and 54.7 grey levels on Sample 6, 3.37 against 28.2 and 31.2 on Sample 15, 5.07 against 5.10 and 5.18 on Sample 14, where 0.1 px hides in 5 levels of noise.
  • Alignment. The two images come together, as the challenge gives them, and get no starting transform: the displacement is the task, most of it below a pixel.
  • Baseline. With no registration the nine test pairs score a TRE of 1.60 px (0.047 to 4.47), rTRE 0.0018: 1.41 px on each of the seven shifts, 0.047 on Sample 14 and 4.47 on Sample 15. The truth is exact, so the room is the whole start.
  • Outside numbers. None compare: the challenge reports bias, noise and spatial resolution of DIC codes on Samples 14 and 15, not point errors.

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
All nine cases are test cases, one a speckle pattern among the challenge's samples whose displacement is documented: fewer than ten, so none is left for development.
Truth
landmarks a known deformation the displacement each moved image was made with, as the challenge documents it (a uniform shift, a sine of rising frequency, a strain concentration with a step), at a 16 x 16 grid of pixels of the reference: 256 a pair
Licence
none stated