PINSIR, or Person Identification Network Stack for Identity Recognition, is a scalable open source end to end solution for face detection and identity recognition.

It is divided into 4 main services Detection, Extraction, Embedding and Comparison with a 5th optional gateway service.

How does it work

Identity recognition requires different steps that need to be framed independently of one another:

• Detection: given an image, for each face present in the image, a bounding box should be output;
• Extraction: each bounding box should be used to extract sub-images containing only the identified faces;
• Embedding: for each sub-image, an appropriate embedding of the face in the image.
• Comparison: given two embedding, a model should be able to say if the embeddings are from the same person or not.
• Gateway: given one or more image, can offer many high level recognition services, such as one to many identity discovery, embedding extraction and so on.

All containers communicate through gRPC generated from proto files available here. While Kotlin/JVM and Java artifacts are available on my Maven repository (instructions below), the Python declarations are not wrapped as a library. To build them copy the proto files and run protoc on your own or clone the project and run ./gradlew :data:generatePythonDefinitions which will be available in data/build/grpc/python.

The training notebooks for the networks inside the Comparison service are available here

Docker images

Images are available for AMD64, ARM32 and ARM64. Working on AMD64 only! Need help to build opencv and pillow in dockerfile, PRs are welcome!

• Detection: lamba92/pinsir.detection:latest
• Extraction: lamba92/pinsir.extraction:latest
• Embedding: lamba92/pinsir.embedding:latest
• Comparison: lamba92/pinsir.comparison:latest
• Gateway: lamba92/pinsir.gateway:latest

Example usage

This is an example compose file for Docker Composer or Docker Swarm:

version: "3.8"
services:
  detection:
    image: lamba92/pinsir.detection:latest

  extraction:
    image: lamba92/pinsir.extraction:latest

  embedding:
    image: lamba92/pinsir.embedding:latest

  comparison:
    image: lamba92/pinsir.comparison:latest
    environment:
      - MODEL: SIMPLE+ #default
#     - MODEL: SIMPLE
#     - MODEL: COMPLEX

  gateway:
    image: lamba92/pinsir.gateway:latest
    ports:
    - 50051:50051
    environment:
      DETECTOR_HOSTNAME: "detection:50051"
      EXTRACTOR_HOSTNAME: "extraction:50051"
      EMBEDDER_HOSTNAME: "embedding:50051"
      COMPARATOR_HOSTNAME: "comparison:50051"

Maven

JVM clients and Kotlin extension function utils available at Maven. Gradle Kotlin DSL:

repositories {
    maven("https://dl.bintray.com/lamba92/com.github.lamba92")
}

dependencies {
    implementation("com.github.lamba92:pinsir-data:LATEST_VERSION")
}

Maven:

<repositories>
    <repository>
      <id>lamba92</id>
      <name>lamba92 repo</name>
      <url>https://dl.bintray.com/lamba92/com.github.lamba92</url>
    </repository>
</repositories>

<dependency>
    <groupId>com.github.lamba92</groupId>
    <artifactId>pinsir-data</artifactId>
    <version>LATEST-VERSION</version>
    <type>pom</type>
</dependency>