COCO-OLAC

COCO Occlusion Labels for All Computer Vision Tasks (COCO-OLAC) provides manually annotated, image-level occlusion labels for 35,000 COCO images: 30,000 training images and 5,000 validation images. Each image is labelled low, mid, or high. The dataset was introduced in COCO-OLAC: A Benchmark for Occluded Panoptic Segmentation and Image Understanding (ICASSP 2025).

Related resources

Research overview: Occlusion-Aware Visual Understanding. Explore the papers, datasets, model checkpoints, and how they relate.

Resource	What it provides
Cityscapes-OLAC dataset	Occlusion annotations for 3,475 Cityscapes images, following the same annotation protocol.
PEMOLA model checkpoints	Occlusion-aware panoptic segmentation models and baselines evaluated on COCO-OLAC and Cityscapes-OLAC, plus an occlusion classifier.

• Paper: https://huggingface.co/papers/2409.12760
• Code and data preparation: https://github.com/wenbo-wei/COCO-OLAC
• Original annotation release: https://github.com/wenbo-wei/COCO-OLAC/releases/tag/annotations-v1.0

This repository contains the additional occlusion annotations. Obtain the original COCO images and semantic, instance, or panoptic annotations separately from the official COCO website.

Files and statistics

The five JSON files are preserved byte-for-byte from the original GitHub release. SHA256SUMS records their SHA-256 checksums.

File	Images	Low	Mid	High	Size (bytes)
occlusion_label_train.json	30,000	6,668	11,251	12,081	822,084
occlusion_label_val.json	5,000	1,134	2,075	1,791	116,791
occlusion_label_val_low.json	1,134	1,134	0	0	30,620
occlusion_label_val_mid.json	2,075	0	2,075	0	56,027
occlusion_label_val_high.json	1,791	0	0	1,791	50,150

The training images come from COCO train2017; validation covers COCO val2017. The three validation subset files partition the same 5,000 validation images and do not add further samples. Training and validation image IDs do not overlap.

Annotation format

Each file is a single JSON object mapping a COCO image ID, encoded as a zero-padded 12-digit string, to an occlusion level:

{
  "000000054334": "high"
}

The levels follow the paper's manual annotation protocol. They describe perceived occlusion at the image level; they are not per-object labels, occlusion masks, or amodal segmentation annotations. Preserve the string IDs and leading zeros when reading the files.

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

Download and use

From the COCO-OLAC or PEMOLA source repository root, download the annotation files to the expected directory:

python -m pip install --upgrade huggingface_hub

hf download weiwb/COCO-OLAC \
  --repo-type dataset \
  --include "occlusion_label_*.json" \
  --local-dir datasets/data/coco_olac

To download and read one split in Python:

import json

from huggingface_hub import hf_hub_download

annotation_path = hf_hub_download(
    repo_id="weiwb/COCO-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)

print(occlusion_labels["000000054334"])  # high
# Corresponding original image: train2017/000000054334.jpg

Follow the COCO-OLAC data preparation instructions or the PEMOLA data preparation instructions for the required image layout and segmentation preprocessing.

Scope and limitations

These labels support occlusion classification and analysis of segmentation performance across occlusion levels. Image-level labels do not measure the visible or occluded area of individual objects. Class frequencies differ, so consider the distribution when interpreting aggregate results. This mirror contains the five annotation files in the linked release; it does not add separate classifier splits or test labels.

License and original data

The source repository releases the additional occlusion-level annotations under Creative Commons Attribution 4.0 International (CC BY 4.0). Please credit the dataset authors and cite the paper when using these annotations.

Original COCO images and annotations remain subject to their respective COCO terms of use. This repository does not redistribute or grant rights to the original images or segmentation annotations.

Citation

@inproceedings{wei2025coco,
  title        = {{COCO-OLAC}: A Benchmark for Occluded Panoptic Segmentation and Image Understanding},
  author       = {Wei, Wenbo and Wang, Jun and Bhalerao, Abhir},
  booktitle    = {ICASSP 2025 -- 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages        = {1--5},
  year         = {2025},
  organization = {IEEE}
}

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