Skip to content

tobiaskalb/disjoint_cityscape_splits

Repository files navigation

Training Splits of Disjoint Cityscapes

This repository contains the training splits for the Class Disjoint Cityscapes used in: Continual Learning for Class- and Domain-Incremental Semantic Segmentation and Improving Replay-Based Continual Semantic Segmentation with Smart Data Selection.

Format

The splits for each subset are saved in a json-Files that contain a list of dictionaries with paths to the training images and the corresponding labels. The dictionaries are structured as follows:

{
  "img": <RELATIVE_PATH_TO_IMG>,
  "label": <RELATIVE_PATH_TO_LABEL>,
  "id": <TRAIN_ID>
  "weight": 1
}

Class Mapping

The original class_ids are mapped to new set of train_ids for Disjoint Cityscapes. Please check the adjusted labels.py which is based on the original Cityscapes implementation. A example on how to map Cityscapes class ids to trainIds can be found in the notebook here.

About

Training Splits of Disjoint Cityscapes

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published