4 releases (1 stable)
1.0.0 | Aug 11, 2023 |
---|---|
0.3.0 | Jul 5, 2023 |
0.2.0 | Jun 30, 2023 |
0.1.0 | Jun 28, 2023 |
#866 in Machine learning
1MB
1.5K
SLoC
Rust Faces - Face Detection Models with Rust Interface
This project aims to provide a Rust interface for multiple state-of-the-art face detection models.
Features
- Integration of multiple face detection models;
- Rust interface: The core implementation is written in Rust, leveraging its safety, performance, and concurrency features for preprocessing, non-maxima suppression, and other non-neural-network computationally intensive tasks.
- ONNX Runtime inference provided by Ort.
- Language bindings: The project will provide bindings to Python (✔️), C (⚙️), C++ (⚙️), Java (⚙️), and C# (⚙️), enabling more developers with face technologies.
- Easy integration
Supported Face Detection Models
The project aims to include a selection of popular and high-performing face detection models, such as:
- BlazeFace - BlazeFace640 and BlazeFace320
- MTCNN (Multi-Task Cascaded Convolutional Networks) - Models from timesler/facenet-pytorch
Please note that the availability of specific models may vary depending on the licensing terms and open-source availability of the respective models.
Usage
Linux Requirements:
- libssl-dev
- pkg-config
Ubuntu: $sudo apt install libssl-dev pkg-config
Install the crate:
$ cargo add rust-faces --features viz
use rust_faces::{
viz, BlazeFaceParams, FaceDetection, FaceDetectorBuilder, InferParams, Provider, ToArray3,
ToRgb8,
};
pub fn main() {
let face_detector =
FaceDetectorBuilder::new(FaceDetection::BlazeFace640(BlazeFaceParams::default()))
.download()
.infer_params(InferParams {
provider: Provider::OrtCpu,
intra_threads: Some(5),
..Default::default()
})
.build()
.expect("Fail to load the face detector.");
let image = image::open("tests/data/images/faces.jpg")
.expect("Can't open test image.")
.into_rgb8()
.into_array3();
let faces = face_detector.detect(image.view().into_dyn()).unwrap();
let mut image = image.to_rgb8();
viz::draw_faces(&mut image, faces);
std::fs::create_dir_all("tests/output").expect("Can't create test output dir.");
image
.save("tests/output/should_have_smooth_design.jpg")
.expect("Can't save test image.");
}
Make OnnxRuntime shared library available
[Linux] If necessary, export your library path to the onnx runtime directory.
$ export LD_LIBRARY_PATH=<path to onnx runtime lib directory>:$LD_LIBRARY_PATH
If you still receive the following error message:
PanicException: ort 1.14 is not compatible with the ONNX Runtime binary found at
onnxruntime.dll
; expected GetVersionString to return '1.14.x', but got '1.10.0'
Try to direct set the environment variable ORT_DYLIB_PATH
:
# bash
$ export ORT_DYLIB_PATH="<your onnx runtime dir>/onnxruntime.so"
# Powershell
> $env:ORT_DYLIB_PATH="<your onnx runtime dir>/onnxruntime.dll"
More details on the Ort project.
Python usage
Requirements
- Rust
- Onnx Runtime - Download one of the releases
$ pip install -e "git+https://github.com/rustybuilder/rust-faces.git#egg=py-rust-faces&subdirectory=python" --install-option="--release"
Usage:
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
import py_rust_faces as prf
detector = prf.blazeface(prf.BlazeFace.Net320)
image = np.array(Image.open(args.image))
rects, _confidences, _landmarks = detector.detect(image)
plt.imshow(image)
for rect in rects:
plt.gca().add_patch(
Rectangle(
(rect[0], rect[1]),
rect[2],
rect[3],
edgecolor="red",
facecolor="none",
lw=2,
)
)
plt.show()
Benchmarks
Algorithm | [min, mean, max] |
---|---|
MTCNN (cpu) | [12.561 ms 13.083 ms 13.646 ms] |
BlazeFace 320 (cpu) | [2.7741 ms 2.7817 ms 2.7900 ms] |
BlazeFace 640 (cpu) | [3.3369 ms 3.3430 ms 3.3498 ms] |
*GPU times are ommited because the architectures are too small, causing them to be similar to CPU.
Contributions
Contributions to the project are welcome! If you have suggestions, bug reports, or would like to add support for additional face detection models or programming languages, please feel free to submit a pull request or open an issue.
Backlog: https://github.com/users/rustybuilder/projects/1
License
This project is licensed under the MIT License.
Dependencies
~16–30MB
~445K SLoC