Home

Awesome

BlendPASS - Blending Panoramic Amodal Seamless Segmentation

BlendPASS [PDF]

Blending Panoramic Amodal Seamless Segmentation (BlendPASS) comprises an unlabeled training set of 2,000 panoramic images and a labeled test set of 100 panoramic images. These images are captured from panoramic cameras in driving scenes across 40 cities for the training set and 20 cities for the test set, all at a resolution of 2048x400 pixels.

We provide finely pixel-level annotations for five segmentation tasks (Semantic, Instance, Panoptic, Amodal Instance and Amodal Panoptic) related to OASS, which greatly extends the semantic labels from DensePASS. It is available at Google Drive.

These annotations cover 19 categories that align with the Cityscapes and are further categorized into Stuff (road, sidewalk, building, wall, fence, pole, light, sign, vegetation, terrain, and sky) and Thing (person, rider, car, truck, bus, train, motorcycle, and bicycle).

PersonRiderCarTruckBusTrainMotorcycleBicycleTotal
#Occluded objects18969094218183381286
#Unoccluded objects613128423824271721674
Total802181751804231541102960

visualization

We provide some code to visualize our dataset. We provid some code to Visualize BlendPASS, please refer to the Visualize folder.

🤝 Publication:

Please consider referencing this paper if you use the code or data from our work. Thanks a lot :)

@inproceedings{cao2024oass,
  title={Occlusion-Aware Seamless Segmentation},
  author={Yihong Cao and Jiaming Zhang and Hao Shi and Kunyu Peng and Yuhongxuan Zhang and Hui Zhang and Rainer Stiefelhagen and Kailun Yang},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024}
}