search

SysCV / shift-dev

SHIFT Dataset DevKit - CVPR2022
https://www.vis.xyz/shift
MIT License
103 stars 10 forks source link

Semantic Segmantion Classes #4

Closed YanivKatz closed 2 years ago

YanivKatz commented 2 years ago

Can you please add a txt/yaml file containing the classes used for semantic segmentation and their indices?

waychin-weiqin commented 2 years ago

Hi @YanivKatz, as the authors generated this dataset using Carla, I managed to find all the information about the classes and their indices for semantic segmentation from the Carla documentation page.

mattiasegu commented 2 years ago

Thanks @waychin-weiqin for replying first! That is indeed correct.

To make it easy for everyone, we will also provide a file with the classes and indices used for semantic segmentation.

Leaving this issue open until then.

suniique commented 2 years ago

Thanks @waychin-weiqin! Yes, we are following Carla's definition for semantic labels. For completeness, here is a table showing the 23 classes and their correspondence to Cityscapes that we used in our experiments. We will update our website to include evaluation details for semantic segmentation.

id name color cityscapes_id_equivalent cityscapes_ignore_in_eval
0 unlabeled ( 0, 0, 0) 0 true
1 building ( 70, 70, 70) 11 false
2 fence (100, 40, 40) 13 false
3 other ( 55, 90, 80) 0 true
4 pedestrian (220, 20, 60) 24 false
5 pole (153, 153, 153) 17 false
6 road line (157, 234, 50) 7 false
7 road (128, 64, 128) 7 false
8 sidewalk (244, 35, 232) 8 false
9 vegetation (107, 142, 35) 21 false
10 vehicle ( 0, 0, 142) 26 false
11 wall (102, 102, 156) 12 false
12 traffic sign (220, 220, 0) 20 false
13 sky ( 70, 130, 180) 23 false
14 ground ( 81, 0, 81) 6 true
15 bridge (150, 100, 100) 15 true
16 rail track (230, 150, 140) 10 true
17 guard rail (180, 165, 180) 14 true
18 traffic light (250, 170, 30) 19 false
19 static (110, 190, 160) 4 true
20 dynamic (170, 120, 50) 5 true
21 water ( 45, 60, 150) 0 true
22 terrain (145, 170, 100) 22 false
HRHLALALA commented 1 year ago

Hi @suniique , I found that some videos have unexpected pixel id of 23. Can you help double check that ? Thanks

In 0ec5-f0e2.mp4:

>>> np.unique(frame)
array([ 0,  1,  3,  4,  7,  9, 10, 11, 13, 19, 20, 23], dtype=uint8)
suniique commented 1 year ago

hey @HRHLALALA, thanks for the question! Yes, the decompressing of video sequences does require many environment settings, otherwise, you may not get exactly the same results. To simply that, we have released all the sematic segmentation labels in the plan zip files of PNGs.

You can download them on our server, for example, the semseg.zip under https://dl.cv.ethz.ch/shift/continuous/videos/1x/train/front/