Pigo is a pure Go face detection, pupil/eyes localization and facial landmark points detection library based on the Pixel Intensity Comparison-based Object detection paper.
| Rectangle face marker | Circle face marker |:--:|:--: | !rectangle | !circle |
Motivation
The reason why this library has been developed is because all of the currently existing solutions for face and facial landmark points detections the Go ecosystem are purely bindings, wrappers to other C/C++ libraries likeOpenCV or dlib, but calling a C program through cgo introduces unnecessary latencies and produces a significant trade-off in terms of performance. Also, in many cases installing OpenCV on various platforms is cumbersome.
Key features
- [x] Does not require OpenCV or any 3rd party modules to be installed
- [x] High processing speed
- [x] There is no need for image preprocessing prior to detection
- [x] There is no need for the computation of integral images, image pyramid, HOG pyramid or any other similar data structure
- [x] The face detection is based on pixel intensity comparison encoded in the binary file tree structure
- [x] Fast detection of in-plane rotated faces
- [x] The cascade files are directly embedded into the generated binary file
- [x] Pupils/eyes localization
- [x] Facial landmark points detection
- [x] Webassembly support 🎉
Todo
- [ ] Object detection and description
The library can also detect in plane rotated faces. For this reason a new -angle parameter has been included into the command line utility. The command below will generate the following result (see the table below for all the supported options).
$ pigo -in input.jpg -out output.jpg -angle=0.8 -iou=0.01
| Input file | Output file |:--:|:--: | !input | !output |
Note: In case of in plane rotated faces the angle value should be adapted to the provided image.
Pupils / eyes localization
Starting from v1.2.0 Pigo supports pupils/eyes localization capabilities also. The implementation is based on Eye pupil localization with an ensemble of randomized trees.Check out this real time demo: https://github.com/esimov/pigo/tree/master/examples/puploc
Facial landmark points detection
v1.3.0 marks a new milestone in the library evolution, Pigo being able to detect facial landmark points. The implementation is based on Fast Localization of Facial Landmark Points.Check out this real time demo: https://github.com/esimov/pigo/tree/master/examples/facial_landmark
WASM (Webassembly) support 🎉
Starting from version v1.4.0 the library has been ported to WASM. This proves the library's real time face detection capabilities, constantly producing ~60 FPS.
WASM demo
To run thewasm demo select the wasm folder and type make. This will automatically open a new tab in the web browser and start capturing the detected faces.
For more details check this subpage: https://github.com/esimov/pigo/tree/master/wasm.
For more awesome WASM demos using the Pigo library check this repo: https://github.com/esimov/pigo-wasm-demos.
Install
Install Go, set yourGOPATH, and make sure $GOPATH/bin is on your PATH.
$ go install github.com/esimov/pigo/cmd/pigo@latest
API
Below is a minimal example of using the face detection API.Prior invoking the cascade classifier function the source image should be converted to grayscale mode, which returns a slice containing the row, column, the scale and the detection score.
src, err := pigo.GetImage("/path/to/image")
if err != nil {
log.Fatalf("Cannot open the image file: %v", err)
}
pixels := pigo.RgbToGrayscale(src)
cols, rows := src.Bounds().Max.X, src.Bounds().Max.Y
cParams := pigo.CascadeParams{
MinSize: 20,
MaxSize: 1000,
ShiftFactor: 0.1,
ScaleFactor: 1.1,
ImageParams: pigo.ImageParams{
Pixels: pixels,
Rows: rows,
Cols: cols,
Dim: cols,
},
}
p := pigo.NewPigo()
// Unpack the binary file. This will return the number of cascade trees,
// the tree depth, the threshold and the prediction from tree's leaf nodes.
classifier := p.Unpack()
angle := 0.0 // cascade rotation angle. 0.0 is 0 radians and 1.0 is 2*pi radians
// Run the classifier over the obtained leaf nodes and return the detection results.
// The result contains quadruplets representing the row, column, scale and detection score.
dets := classifier.RunCascade(cParams, angle)
// Calculate the intersection over union (IoU) of two clusters.
dets = classifier.ClusterDetections(dets, 0.2)
Usage
A command line utility is bundled into the library.$ pigo -in input.jpg -out out.jpg
Supported flags:
$ pigo --help
┌─┐┬┌─┐┌─┐
├─┘││ ┬│ │
┴ ┴└─┘└─┘
Go (Golang) Face detection library.
Version: 1.5.0
-angle float
0.0 is 0 radians and 1.0 is 2*pi radians
-flploc
Activate facial landmark points localization
-in string
Source image (default "-")
-iou float
Intersection over union (IoU) threshold (default 0.15)
-json string
Output the detection points into a json file
-mark
Mark detected eyes (default true)
-marker string
Detection marker: rect|circle|ellipse (default "rect")
-max int
Maximum size of face (default 1000)
-min int
Minimum size of face (default 20)
-out string
Destination image (default "-")
-puploc
Activate pupils/eyes localization
-scale float
Scale detection window by percentage (default 1.15)
-shift float
Shift detection window by percentage (default 0.15)
Notice: The library automatically embeds the cascade files required for pupil/eyes localization and facial landmark points detection.
CLI command examples
You can also use thestdin and stdout pipe commands:
$ cat input/source.jpg | pigo > -in - -out - >out.jpg
in and out default to - so you can also use:
$ cat input/source.jpg | pigo >out.jpg
$ pigo -out out.jpg < input/source.jpg
Using the empty string as value for the -out flag will skip the image generation part. This, combined with the -json flag will encode the detection results into the specified json file. You can also use the pipe - value combined with the -json flag to output the detection coordinates to the standard (stdout) output.
Real time face detection (running as a shared object)
The demos from the examples folder are meant to show how the Pigo library can be integrated into other languages, like in our case Python.
I choose Python just because it's very easy to write a simple script to access the webcam.
How does it works?
From the Python script we access the webcam, converts the pixel data to a byte array and throughcgo we transfer the results as a shared object to the Go program where the core face detection is happening. But as you can imagine this operation is not cost effective, resulting in lower frame rates and higher latency.
Releases
In case you don't have it installed or do not want to install Go, you can obtain the binary file from the releases folder.The library can be accessed as a snapcraft function too.
Benchmark results
Below are the benchmark results obtained running Pigo against GoCV using the same conditions.BenchmarkGoCV-4 3 414122553 ns/op 704 B/op 1 allocs/op
BenchmarkPIGO-4 10 173664832 ns/op 0 B/op 0 allocs/op
PASS
ok github.com/esimov/gocv-test 4.530s
The code used for the above test can be found under the following link: https://github.com/esimov/pigo-gocv-benchmark
Author
- Endre Simo (@simo_endre)
License
Copyright © 2019 Endre SimoThis software is distributed under the MIT license. See the LICENSE file for the full license text.

