This page is provided for historical purposes

Below are download links for EVIMO2 version 1. New projects should use EVIMO2 version 2 for several reasons.

Download EVIMO2v1

Please refer to our documentation for detailed instructions on how to read and parse the data.

The authors would appreciate if you cite our papers:

  1. EV-IMO: Motion Segmentation Dataset and Learning Pipeline for Event Cameras [bibtex]
  2. Learning Visual Motion Segmentation Using Event Surfaces [bibtex]
  3. To be released ArXiV report on EVIMO2

The EVIMO2 Dataset

The Recording Setup

Prophesee Cameras (×2)

  • Resolution: 640×480
  • FOV: ~70°

Samsung DVS Gen3

  • Resolution: 640×480
  • FOV: ~75°

Flea3 (RGB)

  • Resolution: 2080×1552
  • FOV: ~65°

Prophesee Atis Gen3 (not used)

A picture of the rig with two prophese cameras on the sides, a samsung camera in the middle and a flea on the top

Videos

Videos for each sequence are available here

Sequences for Motion Segmentation

A tabletop scenario (shot with all 4 cameras up-close) with up to 20 objects moving independently from the camera. Pixelwise masks, depth maps and trajectories are provided. This dataset is also suitable for object recognition tasks. The ROS bags are here .

Camera Download Links Preview (hover for animation)
flea3_7

  • data .npz

  • data .txt

    • left_camera

  • data .npz

  • data .txt

    • right_camera

  • data .npz

  • data .txt

    • samsung_mono

  • data .npz

  • data .txt

    • Motion Segmentation in Low Light

    These sequences were shot in low light - they feature high amounts of noise for event cameras and can be extremely challenging. The ROS bags are here

    Camera Download Links Preview (Hover for animation)
    left_camera

  • data .npz

  • data .txt

    • right_camera

  • data .npz

  • data .txt

    • samsung_mono

  • data .npz

  • data .txt

    • Structure from Motion / Object Recognition

    The sequences feature a variety of objects scattered on the tabletop, but the scene is rigid (no independent motion). The objects provide depth variation and some degree of diversity from recording to recording, and additionally allow to evaluate object recognition algorithms (pixelwise masks are provided). The ROS bags are here

    This Camera Download Links Preview (Hover for animation)
    flea3_7

  • data .npz

  • data .txt

    • left_camera

  • data .npz

  • data .txt

    • right_camera

  • data .npz

  • data .txt

    • samsung_mono

  • data .npz

  • data .txt

    • Structure from Motion in Low Light

    These sequences were shot in low light - they feature high amounts of noise for event cameras and can be extremely challenging. The ROS bags are here

    Camera Download Links Preview (Hover for animation)
    flea3_7

  • data .npz

  • data .txt

    • left_camera

  • data .npz

  • data .txt

    • right_camera

  • data .npz

  • data .txt

    • samsung_mono

  • data .npz

  • data .txt

    • Benchmarking and Sanity Check Recordings

    The following recordings contain primitive scenarios with rudimentary motions and are provided primarily for sanity testing, benchmarking and debugging of software.

    Sanity Check Recordings In Regular Lighting

    These are sequences shot in a well lit scenes. The ROS bags can be found here .
    Camera Download Links Preview (Hover for animation)
    flea3_7

  • data .npz

  • data .txt

    • left_camera

  • data .npz

  • data .txt

    • right_camera

  • data .npz

  • data .txt

    • samsung_mono

  • data .npz

  • data .text

    • Sanity Check Sequences in Low Light

    These are sequences that were shot in low light. The ROS bags for this can be downloaded here. here

    Camera Download Links Preview (Hover for animation)
    flea3_7

  • data .npz

  • data .txt

    • left_camera

  • data .npz

  • data .txt

    • right_camera

  • data .npz

  • data .txt

    • samsung_mono

  • data .npz

  • data .txt