Skip to content

Latest commit

 

History

History
72 lines (47 loc) · 3.15 KB

datasetInfo.md

File metadata and controls

72 lines (47 loc) · 3.15 KB

Dataset construction

The SeRM dataset consists of a total of 3 parts.

  1. RGB images and semantic labeled images for training (0. SeRM_image)
  2. RGB images for validatation (1. SeRM_image_test)
  3. Localization and Mapping dataset (2. SeRM_mapping)

The structure of each part is as shown in the figure below.

'0. SeRM_image' provides RGB image of 320 x 1280 size and semantic segmentation label, and '1. SeRM_image_test' provides RGB images only. Finally, '2. SeRM_mapping' provides odometry information of the vehicle and the ground truth obtained by GPS for localization & mapping. In addition, RGB images acquired with the camera mounted on the vehicle are provided in 672 x 1280 size. To apply the trained network from 1.SeRM_image, one can use the lower part (320 x 1280) of the 0.original_full image given in 3.SeRM_mapping for inference.

Colormap

Class information

Class information

trainannot label information

Name Train ID category color
Background 0 ignore [0 0 0];
Slow down 1 SRM [0 0 255];
Go ahead 2 SRM [0 128 255];
Trun right 3 SRM [255 128 0];
Turn left 4 SRM [255 192 64];
Ahead or turn right 5 SRM [64 0 255];
Ahead or turn left 6 SRM [128 0 255];
Crosswalk 7 SRM [255 0 255 ];
Double line (yellow) 8 RL [0 255 0];
Double line (blue) 9 RL [192 255 128];
Broken line (white) 10 RL [210 210 210];
Single line (yellow) 11 RL [255 255 0];
Single line (white) 12 RL [128 128 128];
Stop line 13 RL [255 0 0];
Numbers 14 SRM [64 192 64];
Texts 15 SRM [128 192 128];
Others 16 SRM [128 64 80];

Dataset Details

The odometry data (Log_odom.txt) and ground truth (Log_groundtruth.txt) provided by ‘2.SeRM_mapping’ follow the be configuration, respectively.

Log odom.txt

image index, x coordinate, y coordinate, heading angle (radian)

Log groundtruth.txt

image index, x coordinate, y coordinate, heading angle (radian)

Image index shows an index of RGB image in '0.original_full' that matches each x, y, heading angle. The (x, y) coordinate values and heading angle follow the coordinate system below.

Dataset Validation

Please email Wonje Jang your semantic segmentation result of (1. SeRM_image_test) to obtain validation result. The validation results provided are 'recall', 'precision', 'F1-score', 'mIoU', 'mIoU of each class' value as in our paper.