Multiclass Semantic Segmentation Duckietown Dataset

A dataset of multiclass semantic segmentation image annotations for the first 250 images of the "Duckietown Object Detection Dataset".

Raw Image	Segmentated Image

Semantic Classes

This dataset defines 8 semantic classes (7 distinct classes + implicit background class):

Class	XML Label	Description	Color (RGB)
Ego Lane	`Ego Lane`	The lane the agent is supposed to be driving in (default right-hand traffic assumed)	`[102,255,102]`
Opposite Lane	`Opposite Lane`	The lane opposite to the one the agent is supposed to be driving in (default right-hand traffic assumed)	`[245,147,49]`
Road End	`Road End`	Perpendicular red indicator found in Duckietown indicating the end of the road or the beginning of an intersection	`[184,61,245]`
Intersection	`Intersection`	Road tile with no lane markings that has either 3 (T-intersection) or 4 (X-intersection) adjacent road tiles	`[50,183,250]`
Middle Lane	`Middle Lane`	Broken yellow lane in the middle of the road separating lanes	`[255,255,0]`
Side Lane	`Side Lane`	Solid white lane marking the road boundary	`[255,255,255]`
Background	`Background`	Unclassified	- (implicit class)

Notice:

(1) The color assignment is purely a suggestion as the color information encoded in the annotation file is not used by the cvat_preprocessor.py and can therefore be overwritten by any other mapping. The specified color mapping is mentioned here for explanatory and consistency reasons as this mapping is used in dataloader.py (see Usage for more information).

(2) [Ego Lane, Opposite Lane, Intersection] are three semantic classes for essentially the same road tiles - the three classes were added to introduce more information for some use cases. Keep in mind, that some semantic segmentation neural network have a hard time learning the difference between these classes, leading to a poor performance on detecting these classes. In such case, treating these three classes as one "Road" class helps improving the segmentation performance.

(3) The Middle Lane and Side Lane classes were added later and thus only the first 125 images were annotated. If you want to use these, use the segmentation_annotation.xml annotation file. Otherwise, segmentation_annotation_old.xml stores 250 images (including the 125 images from the other annotation file) but without these two classes.

(4) Background is a special semantic class as it is not stored in the annotation file. This class is assigned to all pixels that don't have any other class (see dataloader.py for a reference solution for that).

Usage

Due to the rather large size of the original dataset (~750MB), this repository only contains annotations file stored in CVAT for Images 1.1 format as well as two python files:

cvat_preprocessor.py: A collection of helper functions to read the annotations file and extract the annotation masks stored as polygons.
dataloader.py: A PyTorch-specific example implementation of a wrapper-dataset to use with PyTorch machine learning models.

hamnaanaa / multiclass-semantic-segmentation-duckietown-dataset Goto Github PK