fidopt Interactive view
Plants (2D), same plant over days

MSU-PID

Chlorophyll fluorescence images of one Arabidopsis plant seen from above at 9 am on day 7 (fixed) and day 9 (moving), every leaf outlined and its two tips marked on both (MSU-PID, Michigan State University).

10 test pairsplantsame patientfixed CHLOROPHYLL, moving CHLOROPHYLL2D, distances in pxdelivered alignedresearch only
Look at the images
Read before comparing (9)
  • Plants. Arabidopsis grows its leaves in a rosette that is nearly flat seen from above. Sixteen plants grew in one tray and were photographed from above every hour of light for nine days by fixed cameras; the database cuts each plant out with one box for all its frames.
  • Images. Chlorophyll glows in the near infrared, outlining the living leaf area against a dark background, with the pot's rim faint. The 8-bit grey PNGs are kept as delivered and carry no pixel size: every distance is in pixels, and the rTRE divides it by the image's diagonal, 339 to 366 px.
  • Truth. The authors' team outlined each leaf with a polygon of about 20 clicked points and clicked its outer and inner tips, then corrected inaccurate labels; the annotators are not named. Where outlines overlap, the leaf drawn last takes the pixels: a map holds each leaf's visible part, and a leaf that a younger one covers loses area.
  • Agreement. Two annotators labelling nine images agree with a symmetric best Dice of 0.81 to 0.90, with no figure for the tips. Between two frames of one day on the development plants the outer tips move a median of 3 px, the inner tips 2.2 px, and the leaves overlap at a Dice of 0.87. The tips are good to about 2 to 3 px, and a Dice of about 0.85 to 0.9 is as far as a perfect field can go.
  • Pair. Set on the development plants before any test pair was built: the same hour on both days, so the leaves' daily movement plays no part; the last day, where outlines agree best; and two days, over which the tips move 17 px on the development plants, about six times their own error.
  • Growth. Between the two days a plant's labelled area grows by ×1.45 to ×1.74 on the test pairs, each leaf grows and turns, and up to three new leaves appear: no rigid motion fits. A leaf scores only where both maps hold it: 9 to 11 leaves and 18 to 22 tip pairs a pair.
  • Alignment. Camera and crop boxes are fixed, so the pair is kept as delivered: the background correlates best with no shift on 15 of the 16 plants, within a pixel on plant_12. Unlabelled leaves of neighbouring plants enter some crops at their edges, and leaf 7 of plant_11 reaches its crop's edge on day 9.
  • Baseline. With no registration the 10 test pairs score a leaf Dice of 0.58 (0.53 to 0.69), an HD95 of 23.1 px and a tip error of 16.0 px (14.5 to 17.2), rTRE 0.046. ANTs SyN reaches 0.71 and 11.7 px, greedy 0.72 and 13.4 px, each better than no registration on every pair and without folds.
  • Outside numbers. None: the paper scores leaf segmentation, counting, alignment and tracking, not registration.
AccuracySmoothnessCost
#MethodDiceChangeHD95 pxTRE pxrTRESDlogJFolds %Time sRAM GBGPU GBImages
11–3vismatch matchers · XFeatGPU0.624 ±0.032+0.04018.88 ±1.8714.03 ±0.940.0406 ±0.00280.0440.0032.10.4
21–3vismatch matchers · RoMaGPU0.618 ±0.021+0.03420.55 ±1.0814.70 ±0.590.0425 ±0.00190.0150.0073.16.9
31–3vismatch matchers · LoFTRGPU0.604 ±0.030+0.02021.12 ±2.2915.11 ±1.150.0437 ±0.00290.0240.0032.00.7
44–6vismatch matchers · ALIKED + LightGlueGPU0.454 ±0.145−0.13033.62 ±25.2424.29 ±12.350.0703 ±0.03590.5900.6832.32.1
54–6vismatch matchers · SIFT + LightGlueGPU0.394 ±0.152−0.18941.03 ±26.96147.93 ±278.510.4294 ±0.80881.2593.5732.00.5
64–6vismatch matchers · DISK + LightGlueGPU1 failed0.278 ±0.118−0.30651.18 ±24.02187.86 ±146.900.5468 ±0.43225.67323.0832.13.7
No registration0.584 ±0.03123.10 ±1.0715.96 ±0.680.0462 ±0.00210.0000.00

The order averages each method's ranks on Dice, HD95 px, TRE px, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.

Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. TRE: distance between matching landmarks. rTRE: the TRE over the length of the image's diagonal, to compare datasets of other sizes and units. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.

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
The database's own test split, Arabidopsis plants 7 to 16, one pair a plant; the development cases are its training plants, 1 to 6. The five bean plants are left out: bushy plants whose leaves move fast and in depth, too few to be ranked on their own.
Truth
contours by experts landmarks by experts the leaf outlines the database's authors drew by hand on each image, one label a leaf, numbered as the database follows each leaf over the days; a leaf scores where both images hold it, 9 to 11 a pair the outer and the inner tip of each leaf, clicked by hand on each image by the same authors, paired by the leaf's number: both tips of each leaf both images give, 18 to 22 pairs a pair
Licence
none stated
6 pairs for tuning, never tested

plant_5, plant_4, plant_2, plant_6, plant_3, plant_1