Skip to main content
EURASIP Journal on Image and Video Processing

Table 5 Performance comparison between standard data augmentation and synthetic data usage for the 10-class test set of SegSemOutex

From: Semantic segmentation of textured mosaics

Metric

Class

Case

  

N-N-NoDA\(^{1}\)

N-N-DA\(^{2}\)

S-N-NoDA\(^{3}\)

S-N-DA\(^{4}\)

PA

All

0.9950

0.9940

0.9948

0.9969

mPA

All

0.9950

0.9939

0.9948

0.9969

mIoU

All

0.9900

0.9888

0.9895

0.9938

\(\text {IoU}_1\)

barleyrice002

0.9886

0.9867

0.9692

0.9923

\(\text {IoU}_2\)

canvas005

0.9909

0.9894

0.9940

0.9939

\(\text {IoU}_3\)

canvas022

0.9927

0.9906

0.9957

0.9951

\(\text {IoU}_4\)

carpet005

0.9896

0.9871

0.9718

0.9935

\(\text {IoU}_5\)

chips010

0.9865

0.9846

0.9924

0.9920

\(\text {IoU}_6\)

paper004

0.9926

0.9905

0.9953

0.9949

\(\text {IoU}_7\)

pasta006

0.9877

0.9851

0.9932

0.9930

\(\text {IoU}_8\)

plastic005

0.9874

0.9847

0.9933

0.9929

\(\text {IoU}_9\)

seeds004

0.9894

0.9875

0.9940

0.9938

\(\text {IoU}_{10}\)

wallpaper006

0.9946

0.9936

0.9968

0.9968

  1. Best metric values shown in bold
  2. \(^{1}\): N-N-NoDA: model trained (no data augmentation) and tested on natural (SemSegOutex) textures
  3. \(^{2}\): N-N-DA: model trained (with data augmentation, see Sect. 3.2) and tested on natural (SemSegOutex) textures—same as column N-N in Table 4
  4. \(^{3}\): S-N-NoDA: single model trained (no data augmentation) on synthetic (SemSegOutex-Synth) textures (all seeds combined) and tested on natural (SemSegOutex) textures
  5. \(^{4}\): S-N-DA: single model trained (with data augmentation, see Sect. 3.2) on synthetic (SemSegOutex-Synth) textures (all seeds combined) and tested on natural (SemSegOutex) textures—same as column S-N-All in Table 4
Back to article page