PEMOLA (ICME 2026)

Official checkpoints for Occlusion-Aware Panoptic Segmentation with Joint Position Embedding and Occlusion-Level Attention (ICME 2026).

Related resources

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

Resource	What it provides
COCO-OLAC dataset	Image-level occlusion annotations for 30,000 COCO training images and 5,000 validation images, including validation subsets by occlusion level.
Cityscapes-OLAC dataset	Image-level occlusion annotations for 2,975 Cityscapes training images and 500 validation images, following the same annotation protocol.

Both dataset repositories provide additional occlusion annotations. Obtain the original images and segmentation annotations separately from the official COCO website and the official Cityscapes website.

Download

Download all models from the PEMOLA source repository root:

python -m pip install huggingface_hub
hf download weiwb/PEMOLA --local-dir checkpoints

To download one model, specify its checkpoint filename:

hf download weiwb/PEMOLA mask2former_pemola_coco_olac.pth --local-dir checkpoints

For source code, installation, and inference, see the PEMOLA GitHub repository.

Panoptic Segmentation on COCO-OLAC

Method	Backbone	PQ	PQTh	PQSt	APThpan	mIoUpan	Weights
Mask2Former	ResNet-50	40.7	44.5	35.0	30.0	54.2	download
+ PEMOLA	ResNet-50	41.5	45.2	35.9	30.4	54.8	download
Mask DINO	ResNet-50	44.0	48.5	37.3	33.5	53.4	download
+ PEMOLA	ResNet-50	44.8	49.4	37.8	34.2	55.3	download

Panoptic Segmentation on Cityscapes-OLAC

Method	Backbone	PQ	PQTh	PQSt	APThpan	mIoUpan	Weights
Mask2Former	ResNet-50	61.5	54.0	66.9	35.2	76.1	download
+ PEMOLA	ResNet-50	62.3	55.4	67.2	38.5	77.4	download

Occlusion Classifier

Top-1 accuracies (%) reported in the paper on the COCO-OLAC test split with background-blackened inputs, for occlusion classification.

Backbone	Pretraining	Input	Top-1 Acc (%)	Weights
Swin-L	ImageNet-22K	384	75.3	download