HD1K Benchmark Suite

We are excited to be part of the Robust Vision Challenge 2020. Check out the challenge website and our submission instructions for further details

Welcome to the HD1K Benchmark Suite, an autnonomous driving dataset and benchmark for optical flow. This dataset was created by the Heidelberg Collaboratory for Image Processing in close cooperation with Robert Bosch GmbH.

For the public training dataset, we provide:

  • > 1000 frames at 2560x1080 with diverse lighting and weather scenarios
  • reference data with error bars for optical flow
  • evaluation masks for dynamic objects
  • specific robustness evaluation on challenging scenes

Get the data

Your email will only be used to send you the download links and to keep you informed about updates and bugfixes.

Scene Content

The data was captured in a controlled environment with systematic variation of traffic scenarios, weather, and lighting conditions. The data was acquired at a frame rate of 200Hz with a resolution of 2560x1080.

diverse traffic scenarios

dawn and night scenes

lens flare and reflections

raindrops on windshield

Reference Data with Error Bars

We provide ground truth with uncertainties for optical flow. Creating flow ground truth for real world images is a measurement task and is therefore never perfectly accurate. We propagate measurement uncertainties through the data generation pipeline to obtain per-pixel flow distributions. The interquartile ranges of these distributions are reported as uncertainties.

optical flow ground truth

optical flow uncertainty

Performance Evaluation

We evaluate average performance across all scenes and specific performance for challenging frame types such as darkness or direct sunlight.