khwstolle/csvps

Cityscapes VPS

This dataset is derived from the videos in the validation split of the Cityscapes[^1] dataset. It aggregates the images and metadata from Cityscapes[^1], Cityscapes-VPS[^2] and Cityscapes-DVPS[^3] into a single structured format. This comprehensive derivative was created out of the need for a batteries-included variant of the dataset for academic purposes. Specifically, joining samples from the individual datasets in their original structure (each is organized differently) involves a significant amount of boilerplate code.

This dataset is relevant to computer vision research areas such as:

• Segmentation
• Depth estimation
• Autonomous driving
• Video understanding

Overview

The following variables are included.

1. Images. The input data captured by the left camera from Cityscapes[^1], in 8-bit format. Every sequence has 30 frames.
2. Segmentation labels. Derived from Cityscapes[^1] and Cityscapes-DVPS[^3], these labels provide detailed semantic segmentation and instance segmentation information for 6 frames of every sequence.
3. Depth maps. Improved depth information from Cityscapes-DVPS[^3], offering enhanced quality over the disparity package from Cityscapes[^1], provided for the same samples as the segmentation labels above.
4. Camera calibrations. Includes the intrinsic and extrinsic parameters provided by Cityscapes[^1] for each sequence.
5. Vehicle odometry. Odometry data for each frame, a subset of those provided in Cityscapes[^1].

Files are grouped by split, sequence and frame. This leads to the following structure:

data
    train
        000000
            000000.image.png
            000000.panoptic.png
            000000.depth.png
            000000.vehicle.json
            000000.timestamp.txt
            000000.camera.json
            000001.image.png
            000001.panoptic.png
            000001.depth.png
            000001.vehicle.json
            000001.timestamp.txt
            000001.camera.json
            ...
        000001
            ...
    val
        000000
            ...
        000001
            ....
    test
        000000
            ...
        000001
            ....

The data directory in this repository only contains the segmentation and depth map annotations. The remaining data should be downloaded from official sources using the provided preparation script.

Preparation

1. Install the Cityscapes developer kit using pip.

python -m pip install cityscapesscripts

2. Run the preparation script provided in this repository. Note that this may prompt your Cityscapes account login credentials.

python csvps.py

3. Remove the downloaded Cityscapes archive files to save disk space (optional).

rm -r downloads

Usage

Due to the structure, the dataset can be easily loaded using the webdataset library. To convert the train, val and test directories into a tar archive, run the following command:

for split in train val test; do tar -cf "shards/csvps-$split-000000.tar" "data/$split" done

Subsequently, the dataset can be loaded using the webdataset library:

import webdataset as wds

dataset = wds.WebDataset("shards/csvps-train-{000000}.tar")

License

Please refer to the Cityscapes license for more details.

Citation

If you use this dataset in your research, please cite the original Cityscapes, Cityscapes-VPS, and Cityscapes-DVPS datasets.

[^1]: Cordts et al., “The Cityscapes Dataset for Semantic Urban Scene Understanding” (CVPR 2016)

[^2]: Kim et al., "Video Panoptic Segmentation" (CVPR 2020)

[^3]: Qiao et al., "Learning Visual Perception with Depth-aware Video Panoptic Segmentation" (CVPR 2021)