tobiaskalb / disjoint_cityscape_splits

Training Splits of Disjoint Cityscapes

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.