Cityscapes-OLAC

Cityscapes Occlusion Labels for All Computer Vision Tasks (Cityscapes-OLAC) provides image-level occlusion annotations for the 2,975 training and 500 validation images of Cityscapes. It was introduced in PEMOLA: Occlusion-Aware Panoptic Segmentation with Joint Position Embedding and Occlusion-Level Attention (ICME 2026).

• Paper: https://huggingface.co/papers/2607.18112
• Code and data preparation: https://github.com/wenbo-wei/PEMOLA
• Original annotation release: https://github.com/wenbo-wei/PEMOLA/releases/tag/cityscapes-olac-annotations-v1.0
• Model checkpoints: https://huggingface.co/weiwb/PEMOLA

This repository contains additional occlusion annotations only. It does not contain Cityscapes images or original semantic, instance, or panoptic segmentation annotations. Obtain the original images and gtFine annotations separately from the official Cityscapes website, which requires registration and acceptance of its terms.

Files and format

The original JSON files are preserved without conversion. Each file is a single object mapping a Cityscapes image ID, without the _leftImg8bit.png suffix, to low, mid, or high:

{
  "weimar_000095_000019": "mid",
  "weimar_000140_000019": "mid",
  "weimar_000089_000019": "mid"
}

The tabular dataset viewer is disabled because these files use the original image-ID-to-label mapping format. Download and read them with json.load as shown below.

Labels follow the COCO-OLAC annotation protocol for perceived occlusion at the image level. They are not per-object labels, occlusion masks, or amodal segmentation annotations.

File	Images	Low	Mid	High	Size (bytes)
occlusion_label_train.json	2,975	111	343	2,521	99,241
occlusion_label_val.json	500	19	53	428	16,903
Total	3,475	130	396	2,949	116,144

Download and use

From the PEMOLA repository root, download both JSON files to the expected annotation directory:

python -m pip install --upgrade huggingface_hub

hf download weiwb/Cityscapes-OLAC \
  occlusion_label_train.json occlusion_label_val.json \
  --repo-type dataset \
  --local-dir datasets/data/cityscapes_olac/gtFine

To read the raw annotations in Python:

import json

from huggingface_hub import hf_hub_download

annotation_path = hf_hub_download(
    repo_id="weiwb/Cityscapes-OLAC",
    filename="occlusion_label_train.json",
    repo_type="dataset",
)
with open(annotation_path, encoding="utf-8") as annotation_file:
    occlusion_labels = json.load(annotation_file)

image_id = "weimar_000095_000019"
print(occlusion_labels[image_id])  # mid
# Corresponding original image filename: weimar_000095_000019_leftImg8bit.png

These files support occlusion classification and evaluation of segmentation methods across occlusion levels. Follow the PEMOLA repository's data preparation instructions for the full dataset layout and required preprocessing.

Limitations

The labels describe perceived occlusion for an entire image and do not measure each object's visible or occluded area. Both splits are strongly imbalanced toward high; report class distribution when interpreting aggregate results. This release covers training and validation only and does not provide Cityscapes test labels. Images and segmentation targets are required separately for vision experiments.

License and original data

A separate license for these additional occlusion annotations has not been specified in the original release. The PEMOLA software repository's MIT license is not asserted here as a license for the annotation files.

Cityscapes images and original annotations remain subject to the Cityscapes Terms and Conditions. Those terms permit distribution of additional annotations that do not include or allow reconstruction of the original data, and restrict commercial use of the dataset and derivative works. This repository does not redistribute or grant rights to the original Cityscapes data.

Citation

If you use Cityscapes-OLAC, please cite the paper introducing it:

@inproceedings{wei2026pemola,
  title     = {Occlusion-Aware Panoptic Segmentation with Joint Position Embedding and Occlusion-Level Attention},
  author    = {Wei, Wenbo and Wang, Jun and Raza, Shan and Bhalerao, Abhir},
  booktitle = {IEEE International Conference on Multimedia and Expo (ICME)},
  year      = {2026}
}

When using the original Cityscapes data, also follow its citation requirements.