WildDash Benchmark

Welcome to the WildDash Benchmark. This website provides a dataset and benchmark for semantic and instance segmentation. We aim to improve the expressiveness of performance evaluation for computer vision algorithms in regard to their robustness for driving scenarios under real-world conditions.

Diverse and Challenging Scene Content

We include images from a variety of data sources from all over the world with many different difficult scenarios (e.g. rain, road coverage, darkness, overexposure) and camera characteristics (noise, compression artifacts, distortion). The supplied ground truth format is compatible with Cityscapes.

diverse traffic scenarios

city, highway, and rural locations

scenes from all over the world

poor weather conditions

Focus on Robustness and Performance Evaluation

The main focus of this dataset is testing. It contains data recorded under real world driving situations. Aims of it are:

to compile and provide standard data which can be used for evaluation.
to establish accepted evaluation protocols, data and measures.
to boost the algorithm development on driving applications using computer vision techniques.

The WildDash dataset does not offer enough material to train algorithms by itself. We suggest you use a mixture of material from the Apollo Scape, Berkeley DeepDrive(BDD)/Nexar, Cityscapes, KITTI , and Mapillary datasets for training and the WildDash data for validation and testing.