Note Foundation: image-level occlusion labels for 35,000 COCO images (30,000 training and 5,000 validation), with low, mid and high levels. Defines the annotation protocol reused by Cityscapes-OLAC. Includes five original JSON files and validation subsets by occlusion level. Obtain the original COCO images separately.

Note Extension to urban scenes: applies the COCO-OLAC annotation protocol to 3,475 Cityscapes images (2,975 training and 500 validation). Introduced in the PEMOLA paper to evaluate performance across datasets. Includes occlusion labels; obtain the original Cityscapes images and segmentation annotations separately.

weiwb/PEMOLA

Updated 3 days ago

Note Models: official PEMOLA checkpoints, panoptic segmentation baselines, and an occlusion classifier. PEMOLA uses occlusion cues to improve panoptic segmentation and is evaluated on both COCO-OLAC and Cityscapes-OLAC. Use the two dataset entries for occlusion annotations and the ICME 2026 paper below for the method and experiments.

Note ICASSP 2025. Introduces COCO-OLAC and studies how occlusion affects panoptic segmentation and image understanding. Read this paper for the benchmark and annotation protocol; download its occlusion labels from the COCO-OLAC dataset entry above.

Note ICME 2026. Introduces PEMOLA, combining position embedding modulation with occlusion-level attention, and the Cityscapes-OLAC annotations. Reports evaluations on COCO-OLAC and Cityscapes-OLAC. The PEMOLA model entry contains the checkpoints, and the two dataset entries contain the occlusion annotations.