DVS Benchmark Datasets for Object Tracking, Action Recognition, and Object Recognition

Yuhuang Hu, Hongjie Liu, Michael Pfeiffer and Tobi Delbruck

Institute of Neuroinformatics, University of Zürich and ETH Zürich, Zurich, Switzerland



Quick Fact

Four targeted frame-based datasets
VOT 2015 Dataset Tracking Dataset UCF-50 Dataset Caltech-256 Dataset
Single Target Object Tracking Single Target Object Tracking Action Recognition Object Recognition
Statistics of converted DVS datasets
Name Domain Nr. Recordings Avg. Length/recording (s) Max. FR (keps) Avg. FR (keps)
VOT 2015 Tracking 60 12.25 383.63 251.85
Tracking Dataset Tracking 67 20.70 342.07 197.77
UCF-50 Action Recognition 6676 6.80 238.11 162.62
Caltech-256 Object Recognition 30607 1.01 N/A 110.57

How to Get Datasets


You can directly download the dataset through Zenodo platform. The dataset provides two formats: 1. The raw AEDAT format that is compatible with jAER; and 2. HDF5 format that can be easily used in other programming tools.

HDF5 Format (RECOMMENDED)

Use this link to access the datasets.

jAER AEDAT Format (For inspection)

Use this link to access the datasets.

Updates

2020-06-06: Thanks to a issue report. We found that the basketball sequence in the VOT dataset has incorrect groundtruth labels for the last 46 frames. This issue is caused by the recording itself. We encourage the user to use the first 5768190 events (first 678 frames) for this recording.


Software

We developed a Python package called SpikeFuel that accompanies jAER for

  • Precise control of record logging with Python.
  • User interface for showing video or images in routine.
  • Experiment configuration system (with JSON style).
  • Post signal analysis and selection tools.

A Closer Look



Contacts

Questions about these datasets should be directed to:

References

Acknowledgments

This research is supported by the European Commission project VISUALISE (FP7-ICT-600954), SeeBetter (FP7-ICT-270324), and the Samsung Advanced Institute of Technology.

We gratefully acknowledge the creators of the original datasets.


This dataset is hosted as part of the INI Sensors Group Databases