SpecularDefect9 Dataset


Jingtian Guan   Jingjing Fei   Wei Li   Xiaoke Jiang
Liwei Wu   Yakun Liu   Juntong Xi  
School of Mechanical Engineering, Shanghai Jiao Tong University, Sensetime Research

News

  • 2022-08-09 SpecularDefect9 (classification part) have been released.

Overview Defects

SpecularDefect9 is a large-scale specular defect dataset generated based on 554 float glass with an aluminized front surface and 553 polycarbonate mirrors.
There are a total of 9 classes of defects:
1.Dirt, 2.Fingerprint, 3.Hair, 4.Fiber, 5.Scratch, 6.Staggered_scratch, 7.Scuffing, 8.Concave_defect, 9.Convex_defect.

Classification Task


Specular defect imaging results at each angle include 49 images:
one light intensity image named 'a.png', 24 horizontal fringe images named '1.png'~'24.png', and 24 vertical fringe images named '25.png'~'48.png'. So there are 54,243 images in SpecularDefect9. Each class of defect contains rich identities:

  • 2,173 number of Dirt identities
  • 659 number of Fingerprint identities
  • 664 number of Hair identities
  • 3,157 number of Fiber identities
  • 850 number of Scratch identities
  • 402 number of Staggered_scratch identities
  • 645 number of Scuffing identities
  • 1,324 number of Concave_defect identities
  • 1,306 number of Convex_defect identities
Since each instance contains 49 images, the total number of images is 547,330.
data structures 

	Classification
		├── Test
		│   ├── Concave_defect
		│   │   ├── 1
		│   │   │  │
		│   │   │  ├── a.png
		│   │   │  ├── 1.png
		│   │   │  ├── 2.png
		│   │   │  ├── ...
		│   │   │  ├── ...
		│   │   │  ├── 47.png
		│   │   │  └── 48.png
		│   │   │
		│   │   ├── 2
		│   │   ├── ...
		│   │   ├── ...
		│   │   └── 401
		│   ├── Convex_defect
		│   ├── Dirt
		│   ├── ...
		│   ├── ...
		│   └── Staggered_scratch
		└── Train
			├── Concave_defect
			├── ...
			├── ...
			└── Staggered_scratch

Detection Task

Comming soon.

Downloads

Citation

TBD

Contact


Please contact Jingtian Guan and Wei Li for questions about the dataset.