HDF5 Event File Format

HDF5 is a standard hierarchical self-describing file format ideal to store datasets. Prophesee provides a lossless compression codec for event-based data called ECF (Event Compression Format) that is used in Metavision SDK to support our HDF5 event format.

You can download some HDF5 event files in our Sample Recordings page, or convert one of your RAW file into HDF5 event format using File to HDF5 tool.

Those HDF5 event files can be read with our applications (e.g. Metavision Studio), our C++ samples (e.g. Metavision Viewer) and our Python samples (e.g. Metavision Simple Viewer).

In the C++ API of the SDK, HDF5 event files are handled in Camera class (Metavision::Camera) in the Driver module along with the RAW and DAT file format. Our HDF5 event files can also be directly read and written using Metavision::HDF5EventFileReader and Metavision::HDF5EventFileWriter.

In the Python API of the SDK, HDF5 event files can be read by Core module EventsIterator. At the moment there is no class to write HDF5 event file in Python.

If you want to access the data of our HDF5 event files, you must use our ECF codec. It is delivered with Metavision SDK on both Ubuntu and Windows. It is also available as a shared library and an ECF filter plugin for HDF5 (with self-assigned filter code 0x8ECF) in the hdf5_ecf GitHub repository.


HDF5 event files should not be confused with HDF5 tensor files that can be created with our generate_hdf5 Python script as explained in the tutorial Precomputing features as HDF5 datasets.

Format of HDF5 event files

To read and write HDF5 event files using the SDK, you are not required to know the internal structure of the files as the SDK will handle it for you. But if you want to leverage third-party tools, as introduced in the next section, it is useful to have some information about the format.

Our HDF5 event files is made of:

Each of the two groups has two dataset members:

  • events: a compound dataset made of multiple fields describing the event (encoded with our ECF codec as mentioned above). For example, for CD events, the fields are:

    • x: X coordinate of the event

    • y: Y coordinate of the event

    • p: Polarity of the event

    • t: timestamp (in us) of the event

  • indexes: a compound dataset that contains index information to locate the events in the events dataset. The events are indexed every 2000us. For CD events, the fields are:

    • id: index of the event in the event dataset

    • ts: timestamp of the event of index id

    But note that there is also an attribute called offset on the indexes dataset that contains the value that is added to the timestamp of the event before storing its value in the indexes.

    For example, if we have an HDF5 file with offset of -6384 and indexes starting with those elements:

    • (0, -1)

    • (0, 0)

    • (77012, 2000)

    • (154954, 4000)

    • (234081, 6000)

    Then it should be understood like this:

    • the event of id 77012 has a timestamp of 8384 (2000+6384)

    • the event of id 154954 has a timestamp of 10384 (4000+6384)

    • if you are looking for an event with timestamp ts=11000, then you have to add the offset (-6384) to obtain the timestamp value you should find in the indexes. So in that case you get 4616, hence the event with a timestamp close to 11000 will be found between the events of id 154954 and 234081.

Using HDF5 events files with third-party tool

The HDF5 tools and API can be directly used on our HDF5 event files. The bindings of HDF5 in other languages such as h5py can also be used to manipulate the HDF5 event files produced by Metavision SDK.

HDF5 must be made aware of the ECF codec to correctly interpret the data contained in the files. To do so, the HDF5_PLUGIN_PATH environment variable must point to the location of the ECF codec plugin. This can be accomplished by using the handy setup_env script (if working from the build folder) or modifying your environment (refer to the exact instructions for your platform and software deliverable in the installation guides).


Using custom HDF5 filter plugins is slower than using directly the Metavision SDK. This should be reserved for quick prototyping or applications where the performance is not important.

Example using hdfview

Here, we use hdfview to show the list of attributes of our driving sample HDF5 event file:

HDF5 event file structure

And here, we look at the structure (x, y, p, t) of the events member of the CD dataset:

HDF5 storage of events

Example using h5dump

You can quickly extract information from one of our recording using h5dump. For example, on Ubuntu, we can extract the resolution (via the geometry attribute) of the turning spinner sample recording:

h5dump -a geometry spinner.hdf5

which would give the following ouput :

HDF5 "spinner.hdf5" {
ATTRIBUTE "geometry" {
      CTYPE H5T_C_S1;
  DATA {
  (0): "640x480"

Or display the value of the first 3 events :

h5dump -d "/CD/events" -k 3 spinner.hdf5

which would output :

HDF5 "spinner.hdf5" {
DATASET "/CD/events" {
      H5T_STD_U16LE "x";
      H5T_STD_U16LE "y";
      H5T_STD_I16LE "p";
      H5T_STD_I64LE "t";
  DATASPACE  SIMPLE { ( 54165303 ) / ( H5S_UNLIMITED ) }
      START ( 0 );
      STRIDE ( 1 );
      COUNT ( 1 );
      BLOCK ( 3 );
      DATA {
      (0): {
      (1): {
      (2): {


To use HDF5 tools on Ubuntu, you may need to install an additional package :

sudo apt -y install hdf5-tools

Example using h5py

As a toy example, we want to count how many events occurred between two timestamps in a specified ROI :

import h5py
with h5py.File("spinner.hdf5", "r") as f:
  evts = f["CD"]["events"]
  print(len(evts[(evts['t']<49330) & (evts['t']>48270) & (evts['x']<300) & (evts['y']>300)]))

which would output :



This is a simple and quite inefficient way of extracting information from a HDF5 recording. For better approaches, read the h5py documentation or use the Metavision SDK.