PPS: Wild Panoramic Panoptic Segmentation dataset

Google Drive

Description

We provide a diverse dataset consisting of 80 panoramic images from 40 different cities taken from WildPASS along with panoptic annotations for the most essential street scence classes (Stuff: Road, Sidewalk & Thing: Person, Car) in cityscapes annotation format.

Trained Models

We provide our proposed robust seamless segmentation model. The weights can be downloaded from Google drive Robust Seamless Segmentation Model.

Getting started with the robust seamless segmentation model

Setup

Follow the installation steps for the seamless segmentation model as described in the official repository
Download the provided weights
Download the WildPPS dataset. If you are using a custom dataset, make sure your dataset is in the official cityscapes data format. If it is not the case, use cityscapesScripts to bring it to the required format.

Running inference

Panoptic predictions can be computed as described in the official seamseg repository. Make sure to use the --raw flag to compute quantitative results.
Use the provided robust_seamseg.ini config file and update the path to the weights in line 13

Computing the Panoptic Quality

Convert the provided WildPPS dataset into the required format for the seamless segmentation model with the help of this script or download the precomputed WildPPS_seamsegformat dataset. This is only necessary if you use the seamless segmentation models. Other popular libraries such as DETECTRON2 can directly work with the normal format.
Use the provided measure_panoptic.py script to compute the panoptic score of the predictions. Make sure to have the seamless segmentation model installed.
python measure_panoptic.py --gt <Location of converted WildPPS dataset> --target <Location of seamseg model predictions> --result <Location to write the results to>

Publication

If you use our dataset, model or other code please consider referencing one of our following papers

Jaus, Alexander, Kailun Yang, and Rainer Stiefelhagen. "Panoramic panoptic segmentation: Towards complete surrounding understanding via unsupervised contrastive learning." 2021 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2021. [PDF]

Jaus, Alexander, Kailun Yang, and Rainer Stiefelhagen. "Panoramic panoptic segmentation: Insights into surrounding parsing for mobile agents via unsupervised contrastive learning." IEEE Transactions on Intelligent Transportation Systems (2023). [PDF]

alexanderjaus/PPS