README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: mask
      dtype: image
  splits:
    - name: train
      num_bytes: 547096711
      num_examples: 5000
    - name: validation
      num_bytes: 117520888
      num_examples: 1000
    - name: test
      num_bytes: 496235957.648
      num_examples: 4428
  download_size: 1184441980
  dataset_size: 1160853556.648
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: apache-2.0
task_categories:
  - image-segmentation
size_categories:
  - 1K<n<10K

Links

• Paper: https://arxiv.org/pdf/2101.08461
• Repository: https://github.com/xieenze/Trans2Seg

Dataset composition

There might be some differences with what the paper describes. We try to investigate and get this unique colors in the mask images:

Color (np.uint32(0), np.uint32(0), np.uint32(0)): Found in 10428 images
Color (np.uint32(235), np.uint32(255), np.uint32(7)): Found in 997 images
Color (np.uint32(255), np.uint32(51), np.uint32(7)): Found in 3313 images
Color (np.uint32(120), np.uint32(120), np.uint32(120)): Found in 1572 images
Color (np.uint32(224), np.uint32(5), np.uint32(255)): Found in 3059 images
Color (np.uint32(150), np.uint32(5), np.uint32(61)): Found in 1472 images
Color (np.uint32(204), np.uint32(5), np.uint32(255)): Found in 340 images
Color (np.uint32(4), np.uint32(250), np.uint32(7)): Found in 603 images
Color (np.uint32(204), np.uint32(255), np.uint32(4)): Found in 501 images
Color (np.uint32(140), np.uint32(140), np.uint32(140)): Found in 410 images
Color (np.uint32(120), np.uint32(120), np.uint32(70)): Found in 279 images
Color (np.uint32(6), np.uint32(230), np.uint32(230)): Found in 90 images
Color (np.uint32(255), np.uint32(0), np.uint32(0)): Found in 1 images # red, this is a cup, mapped to (255, 51, 7)

Index	Class Name	Image Count (from paper)	Closest Color Count	Ordered RGB Value
0	background	(not listed)	10428	[0, 0, 0]
1	box	603	603	[4, 250, 7]
2	bottle	1472	1472	[150, 5, 61]
3	window	501	501	[204, 255, 4]
4	eyeglass	410	410	[140, 140, 140]
5	freezer	90	90	[6, 230, 230]
6	jar/kettle	997	997	[235, 255, 7]
7	door	1572	1572	[120, 120, 120]
8	cup	3315	3313+1	[255, 51, 7]
9	wall	3059	3059	[224, 5, 255]
10	bowl	340	340	[204, 5, 255]
11	shelf	280	279	[120, 120, 70]

Split info

We use the same split composition as the original dataset.

License info

We attach Apaceh 2.0 license as it is the license used in the github repository, please look up to the paper and the repository for better details.