AEStream efficiently reads sparse events from an input source and streams it to an output sink. AEStream supports reading from files, USB cameras, as well as network via UDP and can stream events to files, network over UDP, and peripherals such as GPUs and neuromorphic hardware.
AEStream can be installed with the Python package manage pip.
However, AEStream depends on PyTorch, LZ4, and libcaer, so please install these dependencies manually before you install AEStream.
For PyTorch, recall whether you are using the CPU or CUDA version and use the corresponding command below to install AEStream:
| PyTorch version | Command |
|---|---|
| With CPU | pip install aestream --extra-index-url https://download.pytorch.org/whl/cpu |
| With CUDA | pip install aestream --extra-index-url https://download.pytorch.org/whl/cu116 |
Note that this uses CUDA 11.6. Other versions can be found at the website of PyTorch. We do not currently support other platforms than Linux, but contributions are most welcome.
The Python API exposes two classes for reading DVS data sources into PyTorch tensors: USBInput and UDPInput.
# Stream events from a DVS camera over USB
with USBInput((640, 480)) as stream:
while True:
frame = stream.read() # Provides a (640, 480) tensor
...# Stream events from UDP port 3333 (default)
with UDPInput((640, 480), port=3333) as stream:
while True:
frame = stream.read() # Provides a (640, 480) tensor
...More examples can be found in our example folder.
Please note the examples may require additional dependencies (such as Norse for spiking networks or PySDL for rendering). To install all the requirements, simply stand in the aestream root directory and run pip install -r example/requirements.txt
We stream events from a camera connected via USB and process them on a GPU in real-time using the spiking neural network library, Norse using fewer than 50 lines of Python.
The left panel in the video shows the raw signal, while the middle and right panels show horizontal and vertical edge detection respectively.
The full example can be found in example/usb_edgedetection.py
Installing AEStream also gives access to the command-line interface (CLI) aestream.
To use aestraem, simply provide an input source and an optional output sink (defaulting to STDOUT):
aestream input <input source> [output <output sink>]| Input | Description | Usage |
|---|---|---|
| DAVIS, DVXPlorer | Inivation DVS Camera over USB | input inivation |
| EVK Cameras | Prophesee DVS camera over USB | input prophesee |
| File | AEDAT file format as .aedat or .aedat4 |
input file x.aedat4 |
| Output | Description | Usage |
|---|---|---|
| STDOUT | Standard output (default output) | output stdout |
| Ethernet over UDP | Outputs to a given IP and port using the SPIF protocol | output udp 10.0.0.1 1234 |
.aedat4 file |
Output to .aedat4 format |
output file my_file.aedat4 |
| CSV file | Output to comma-separated-value (CSV) file format | output file my_file.txt |
| Example | Syntax |
|---|---|
| Read file to STDOUT | aestream input file example/sample.aedat4 |
| Stream DVS Davis346 (USB 0:2) by iniVation AG to STDOUT (Note, requires Inivation libraries) | aestream input inivation output stdout |
| Stream Prophesee 640x480 (serial Prophesee:hal_plugin_gen31_fx3:00001464) to STDOUT (Note, requires Metavision SDK) | aestream input output stdout |
| Read file to remote IP X.X.X.X | aestream input file example/sample.aedat4 output udp X.X.X.X |
AEStream requires libtorch. Metavision SDK and libcaer are optional dependencies, but are needed for connecting to Prophesee and Inivation cameras respectively.
AEStream is based on C++20. Since C++20 is not yet fully supported by all compilers, we recommend using GCC >= 10.2.
To build the binaries of this repository, run the following code:
export CMAKE_PREFIX_PATH=`absolute path to libtorch/`
# Optional: Ensure paths to libcaer, Metavision, or OpenCV is in place
mkdir build/
cd build/
cmake -GNinja ..
ninja
If your default C++ compiler doesn't support C++ 20, you will have to install an up-to-date compiler and provide the environmental variable CXX.
For instance like this: CXX=/path/to/g++ cmake -GNinja ..
For Inivation cameras, the libcaer library needs to be available, either by a -DCMAKE_PREFIX_PATH flag to cmake or included in the PATH environmental variable.
For examble: cmake -GNinja -DCMAKE_PREFIX_PATH=/path/to/libcaer.
Inivation made the library available for most operating systems, but you may have to build it yourself.
For Prophesee cameras, a version of the Metavision SDK needs to be present.
The open-source version the SDK openeb is available with installation instructions at https://github.com/prophesee-ai/openeb.
Using openeb, it should be sufficient to install it using cmake && make && make install to put it in your path.
Otherwise, you can point to it using the -DCMAKE_PREFIX_PATH option in cmake.
AEStream is created by
- Jens E. Pedersen (@GitHub jegp), doctoral student at KTH Royal Institute of Technology, Sweden.
- Christian Pehle (@GitHub cpehle), PostDoc at University of Heidelberg, Germany.
The work has received funding from the EC Horizon 2020 Framework Programme under Grant Agreements 785907 and 945539 (HBP) and by the Deutsche Forschungsgemeinschaft (DFG, German Research Fundation) under Germany's Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster).
Thanks to Philipp Mondorf for interfacing with Metavision SDK and preliminary network code.
Please cite aestream if you use it in your work:
@misc{aestream,
doi = {10.48550/ARXIV.2212.10719},
url = {https://arxiv.org/abs/2212.10719},
author = {Pedersen, Jens Egholm and Conradt, Jörg},
title = {AEStream: Accelerated event-based processing with coroutines},
publisher = {arXiv},
year = {2022},
}