Awesome
A Python package for decoding RAW and DAT files (Prophesee) to structured NumPy arrays of events.
Supported formats
Installation
You can install the library through pip:
pip install expelliarmus
Thanks to @Tobias-Fischer, the package is also available on conda-forge!
The package is tested on Windows, MacOS and Linux. Join us on Discord to propose features or to signal bugs!
Documentation
Check out readthedocs!
Getting started
expelliarmus is a library that allows to decode binary files generated by Prophesee cameras to NumPy structured arrays.
The expelliarmus API contains a single class called Wizard, that contains many methods to read a file all at once, in chunks of chunk_size events or in time windows of time_window milliseconds. There are also additional methods to save structured NumPy arrays to different Prophesee encoding formats.
Read a file
Let us download this file from the Prophesee website.
from pathlib import Path
import requests
prophesee_url = "https://dataset.prophesee.ai/index.php/s/fB7xvMpE136yakl/download"
fpath = Path("./pedestrians.raw")
# Downloading the file if it is not available.
if not fpath.is_file():
print("Downloading the file...", end=" ")
open(fpath, 'wb').write(requests.get(prophesee_url).content)
print("done!")
else:
print("File already available.")
Downloading the file... done!
The file that we downloaded is an EVT3 one; hence, we need to create a Wizard object choosing an "evt3" encoding.
from expelliarmus import Wizard
wizard = Wizard(encoding="evt3")
The file to be read can be specified in three ways:
- passing the
fpathargument to theWizardconstructor at object creation time. - using the
set_file()method. - passing the file path to the
read()method.
Let us use the second way.
wizard.set_file(fpath)
Now we can use the read() method to read the binary file to a NumPy structured array.
arr = wizard.read()
print(f"First event encoded as (t, x, y, p): {arr[0]}")
print(f"Number of events: {len(arr)}.")
print(f"Recording duration: {(arr[-1]['t']-arr[0]['t'])//int(1e6)} s.")
First event encoded as (t, x, y, p): (5840504, 707, 297, 0)
Number of events: 39297796.
Recording duration: 60 s.
Reading in chunks
The file could be too large to be read all at once in an array; for this reason, expelliarmus provides two generator methods: read_chunk() and read_time_window(), to read a file in chunks of a chunk_size events or in time windows of time_window milliseconds, respectively. Let us start from the first method.
chunk_size = 8192
wizard.set_chunk_size(chunk_size)
# Calling the generator once:
chunk = next(wizard.read_chunk())
print(f"Chunk length: {len(chunk)}.")
print(f"Chunk duration: {(chunk[-1]['t']-chunk[0]['t'])/1e3:.2f} ms.")
print(f"Chunk first event: {chunk[0]}.")
Chunk length: 8192.
Chunk duration: 154.27 ms.
Chunk first event: (5840504, 707, 297, 0).
Let us read a chunk of at most time_window milliseconds duration from the file:
time_window = 5
wizard.set_time_window(time_window)
# Calling the generator once.
chunk = next(wizard.read_time_window())
print(f"Chunk length: {len(chunk)}.")
print(f"Chunk duration: {(chunk[-1]['t']-chunk[0]['t'])/(1e3):.2f} ms.")
print(f"Chunk first event: {chunk[0]}.")
Chunk length: 47.
Chunk duration: 4.50 ms.
Chunk first event: (5840504, 707, 297, 0).
Conversion among file formats
Suppose that you have a really large file encoded in DAT, like this one. You might want to convert it to EVT2 to save disk space and have better read performance. expelliarmus allows you to do that. Let us download the file first.
prophesee_url = "https://dataset.prophesee.ai/index.php/s/YAri3vpPZHhEZfc/download"
fpath = Path("./spinner.dat")
# Downloading the file if it is not available.
if not fpath.is_file():
print("Downloading the file...", end=" ")
open(fpath, 'wb').write(requests.get(prophesee_url).content)
print("done!")
else:
print("File already available.")
Downloading the file... done!
First we change wizard encoding and, then, we read the DAT file to an array.
wizard.set_encoding("dat")
arr = wizard.read(fpath)
print(f"First event encoded as (t, x, y, p): {arr[0]}")
print(f"Number of events: {len(arr)}.")
print(f"Recording duration: {(arr[-1]['t']-arr[0]['t'])/1e6:.2f} s.")
First event encoded as (t, x, y, p): (0, 237, 121, 1)
Number of events: 54165303.
Recording duration: 5.00 s.
Now we define a second Wizard object with EVT2 encoding and we use its save() method to convert the file from DAT to EVT2.
import numpy as np
wizard_evt2 = Wizard(encoding="evt2")
new_fpath = Path("./spinner_evt2.raw")
wizard_evt2.save(fpath=new_fpath, arr=arr)
Let us check that the files are consistent.
new_arr = wizard_evt2.read(new_fpath)
print(f"Durations: DAT = {(arr[-1]['t']-arr[0]['t'])/1e6:.2f} s, \
EVT2 = {(new_arr[-1]['t']-new_arr[0]['t'])/1e6:.2f} s.")
are_equal = True
for coord in ('t', 'x', 'y', 'p'):
are_equal = are_equal and np.equal(arr[coord], new_arr[coord]).all()
print(f"The two arrays are {'not' if (not are_equal) else ''}identical.")
Durations: DAT = 5.00 s, EVT2 = 5.00 s.
The two arrays are identical.
There are other methods available in the API. Check it out!
A small benchmark
Here it is a small benchmark using expelliarmus on the file formats supported. Benchmarking is run on this file, converted from EVT3 to DAT and EVT2 using the save() method. The data shows the file size, read time for the full file and read time for reading the file in chunks and time windows. The performance is compared against HDF5, HDF5 LZF, HDF5 GZIP and NumPy.
Full file read
------------------------------------------------------------------------------------------------------------
Software | Size [MB] | Diff. DAT | Diff. EVT2 | Diff. EVT3 | Time [s] | Diff. DAT | Diff. EVT2 | Diff. EVT3
------------------------------------------------------------------------------------------------------------
exp. DAT | 851 | -0% | +100% | +143% | 1.15 | -0% | +43% | -41%
------------------------------------------------------------------------------------------------------------
exp. EVT2 | 426 | -50% | -0% | +22% | 0.80 | -30% | -0% | -59%
------------------------------------------------------------------------------------------------------------
exp. EVT3 | 350 | -59% | -18% | -0% | 1.95 | +70% | +144% | -0%
------------------------------------------------------------------------------------------------------------
hdf5 | 1701 | +100% | +299% | +386% | 0.73 | -36% | -8% | -62%
------------------------------------------------------------------------------------------------------------
hdf5_lzf | 746 | -12% | +75% | +113% | 3.09 | +170% | +287% | +58%
------------------------------------------------------------------------------------------------------------
hdf5_gzip | 419 | -51% | -2% | +20% | 5.60 | +389% | +600% | +187%
------------------------------------------------------------------------------------------------------------
numpy | 1701 | +100% | +299% | +386% | 0.32 | -72% | -60% | -84%
------------------------------------------------------------------------------------------------------------
Time windowing read
------------------------------------------------------------------------------------------------------------
Software | Size [MB] | Diff. DAT | Diff. EVT2 | Diff. EVT3 | Time [s] | Diff. DAT | Diff. EVT2 | Diff. EVT3
------------------------------------------------------------------------------------------------------------
exp. DAT | 851 | -0% | +100% | +143% | 1.58 | -0% | +4% | -39%
------------------------------------------------------------------------------------------------------------
exp. EVT2 | 426 | -50% | -0% | +22% | 1.51 | -4% | -0% | -42%
------------------------------------------------------------------------------------------------------------
exp. EVT3 | 350 | -59% | -18% | -0% | 2.58 | +64% | +71% | -0%
------------------------------------------------------------------------------------------------------------
hdf5 | 1701 | +100% | +299% | +386% | 1.02 | -35% | -32% | -60%
------------------------------------------------------------------------------------------------------------
hdf5_lzf | 746 | -12% | +75% | +113% | 3.82 | +143% | +153% | +48%
------------------------------------------------------------------------------------------------------------
hdf5_gzip | 419 | -51% | -2% | +20% | 6.88 | +337% | +355% | +166%
------------------------------------------------------------------------------------------------------------
Chunk read
------------------------------------------------------------------------------------------------------------
Software | Size [MB] | Diff. DAT | Diff. EVT2 | Diff. EVT3 | Time [s] | Diff. DAT | Diff. EVT2 | Diff. EVT3
------------------------------------------------------------------------------------------------------------
exp. DAT | 851 | -0% | +100% | +143% | 1.64 | -0% | +3% | -22%
------------------------------------------------------------------------------------------------------------
exp. EVT2 | 426 | -50% | -0% | +22% | 1.58 | -3% | -0% | -24%
------------------------------------------------------------------------------------------------------------
exp. EVT3 | 350 | -59% | -18% | -0% | 2.09 | +28% | +32% | -0%
------------------------------------------------------------------------------------------------------------
hdf5 | 1701 | +100% | +299% | +386% | 4.20 | +157% | +166% | +101%
------------------------------------------------------------------------------------------------------------
hdf5_lzf | 746 | -12% | +75% | +113% | 10.36 | +534% | +555% | +395%
------------------------------------------------------------------------------------------------------------
hdf5_gzip | 419 | -51% | -2% | +20% | 17.23 | +954% | +989% | +724%
------------------------------------------------------------------------------------------------------------
Contributing
Please check our documentation page for more details on contributing.
About
This project has been created by Fabrizio Ottati and Gregor Lenz.