Results 结果

Accuracy of the CrackEvo models on the team's laboratory tests.

Main benchmark: 5-fold cross-validation by test specimen 主结果

Every labelled test sequence is assigned to one fold and tested exactly once (45 sequences, 832 frames). Frames of one specimen never appear in both training and test. Values are mean ± standard deviation over folds.

model	pixel IoU	pixel F1	tolerant F1 (2 px)	clDice	sequence-mean IoU
UNet-ResNet34 + auxiliary public data	0.503 ± 0.078	0.666 ± 0.069	0.826 ± 0.022	0.699 ± 0.022	0.546 ± 0.045
UNet-ResNet34, own data only	0.486 ± 0.064	0.652 ± 0.058	0.832 ± 0.022	0.695 ± 0.016	0.559 ± 0.015

By test family

family	pixel IoU	tolerant F1	clDice
concrete, uniaxial	0.641 ± 0.117	0.882 ± 0.115	0.822 ± 0.125
mortar, uniaxial	0.438 ± 0.111	0.728 ± 0.093	0.614 ± 0.061
SHPB discs, coal and sandstone	0.562 ± 0.060	0.863 ± 0.018	0.714 ± 0.019

Early vs late frames

Early = crack area below half of the maximum of its sequence.

stage	frames	tolerant precision	tolerant recall	tolerant F1
early	461	0.812	0.816	0.790
late	371	0.896	0.862	0.868

Evolution curves follow the hand labels 演化曲线与人工标注一致

Pixel scores are strict for thin, hand-drawn cracks, but crack evolution is read from curves. For every held-out test we compared the crack area predicted by the model with the hand-labelled area over the labelled frames: rank correlation over time, and the mean difference of the growth curves normalised from 0 (first labelled frame) to 1 (last).

family	tests	rank correlation (median)	growth-curve difference (median)
concrete, uniaxial	4	0.973	0.064
mortar, uniaxial	9	0.983	0.075
SHPB discs, coal and sandstone	32	0.998	0.071

Single split (protocol A) 单次划分

model	pixel IoU	tolerant F1	clDice	test frames / sequences
UNet-ResNet34 (the model used on this site)	0.453	0.834	0.690	205 / 10

Image scale matters 尺度的影响

Tolerant F1 of the same model when images are shrunk before inference (mortar family, whose cracks are about 1.7 px wide). Use the original resolution for thin cracks.

longest side	tolerant F1
original	0.747
1536 px	0.693
1024 px	0.643
768 px	0.595

Protocol 评测协议

Splits are made by test specimen, never by frame: neighbouring frames of one video are almost identical and would leak.
Pixel IoU/F1 are strict; the tolerant F1 counts a pixel as correct within 2 px, because hand-drawn crack strokes are 1–4 px off the visible crack in about a fifth of the pixels; clDice measures topology on skeletons.
Public crack datasets are used for training only and never for testing.