# Cityscapes

Cityscapes is a large-scale dataset for semantic urban scene understanding, containing diverse street scenes from 50 cities with pixel-level annotations for training and evaluating computer vision models.

Cityscapes is a large-scale dataset designed for semantic understanding of urban street scenes. It provides a diverse set of stereo video sequences recorded from a moving vehicle in 50 cities across Germany and neighboring countries, with high-quality pixel-level annotations for 5,000 frames and coarser polygonal annotations for an additional 20,000 frames. The dataset was introduced in 2016 by researchers at Daimler AG, TU Darmstadt, and other institutions, and has become a standard benchmark for tasks such as semantic segmentation, instance segmentation, and panoptic segmentation in autonomous driving and [computer-vision](https://www.wikiprompt.org/wiki/computer-vision) research.

The primary goal of Cityscapes is to enable the development and evaluation of algorithms that can understand the visual environment of a vehicle in real-world conditions. The dataset captures the complexity of urban driving, including varying weather conditions, seasons, and traffic situations, making it a challenging and realistic testbed. Its annotations cover 30 classes grouped into eight categories: flat surfaces (roads, sidewalks), humans (pedestrians, riders), vehicles (cars, trucks, buses), construction (buildings, walls), objects (traffic signs, poles), nature (vegetation, terrain), sky, and void (unlabeled areas). This fine-grained taxonomy supports detailed scene analysis and is widely used for training [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) models in autonomous-driving systems.

## Data Collection and Annotation

The dataset was captured using a stereo camera rig mounted on a car, recording at 17 frames per second with a resolution of 2048x1024 pixels. The recording sessions took place in spring, summer, and fall, covering sunny, cloudy, and rainy conditions, but not winter or night scenes. The 5,000 finely annotated frames were selected from 27 cities, while the 20,000 coarsely annotated frames come from an additional 23 cities, ensuring geographic diversity. Annotation was performed by a professional team, with each fine image taking about 90 minutes to label, using a polygonal tool that allows precise boundary delineation. The coarse annotations, which take about 7 minutes per image, use polygons with fewer vertices and are intended for pre-training or weakly supervised learning.

## Benchmark Tasks and Metrics

The dataset is primarily used for semantic segmentation, where each pixel is assigned a class label. The standard evaluation metric is the mean Intersection-over-Union (mIoU) across all classes, with separate reporting for the 19 classes that are used for training and evaluation (the 'void' class is excluded). Cityscapes also supports instance segmentation (distinguishing individual objects within a class, such as separate cars) and panoptic segmentation (combining semantic and instance segmentation). The benchmark has a public leaderboard where researchers can submit results, and it has driven significant progress in segmentation models, from early fully-convolutional-networks to modern [transformer](https://www.wikiprompt.org/wiki/transformer)-based architectures like segformer and mask2former.

## Impact on Computer Vision

Cityscapes has become one of the most influential datasets in computer vision, with thousands of citations and widespread use in both academia and industry. It has spurred the development of domain adaptation techniques, since models trained on Cityscapes often need to generalize to other urban environments, such as those in [waymo](https://www.wikiprompt.org/wiki/waymo) or [tesla-autopilot](https://www.wikiprompt.org/wiki/tesla-autopilot) data. The dataset also popularized the use of synthetic data, with tools like gta-v and carla used to generate additional training data that is then fine-tuned on Cityscapes. Its annotations have been used to train models for [semantic-segmentation](https://www.wikiprompt.org/wiki/semantic-segmentation), depth-estimation, and [object-detection](https://www.wikiprompt.org/wiki/object-detection), and it remains a standard evaluation suite for new architectures, including those based on [neural-network](https://www.wikiprompt.org/wiki/neural-network) and [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) paradigms.

## Limitations and Extensions

Despite its success, Cityscapes has known limitations. It covers only European cities, mostly in Germany, which introduces a bias toward European road layouts, signage, and vehicle types. The dataset does not include night scenes, heavy rain, or snow, limiting its applicability to adverse weather conditions. To address these gaps, researchers have created extensions and complementary datasets, such as cityscapes-dark for night scenes and acdc (Adverse Conditions Dataset) for fog, rain, and snow. Additionally, the coarse annotations are less precise, which can affect models trained on them. Nevertheless, Cityscapes remains a foundational resource, and its benchmark continues to be actively used for evaluating state-of-the-art models in urban scene understanding.

## See Also

- [semantic-segmentation](https://www.wikiprompt.org/wiki/semantic-segmentation)
- autonomous-driving
- [computer-vision](https://www.wikiprompt.org/wiki/computer-vision)
- dataset

---
Source: https://www.wikiprompt.org/wiki/cityscapes
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-07T02:34:23.178922+00:00