CompreFace Documentation

The frontend follows a structured Angular architecture organized by responsibility:

Folder Structure

• styles/: Common SCSS styles.
• core/: Global framework-based services.
• data/: Global models, enums, classes, and interfaces (e.g., user models, permissions, roles). Note: Do not place DTOs here.
• feature/containers/: Smart/container components that communicate with the store via facades.
• feature/components/: Presentational/dumb components that communicate with containers via @Input and @Output.
• pages/: Router modules (e.g., home, login, 404) and components responsible for page layout and feature composition.

State Management (Store)

State is managed using NgRx. The store/ directory contains the root index.ts with the union store and reducers. Feature-specific state is organized as follows:

• store/featureName/module: Encapsulated feature store using StoreModule.forFeature.
• store/featureName/actions: Actions handled by effects or reducers; can be called via facades.
• store/featureName/effects: Side-effect handlers (e.g., API calls).
• store/featureName/selectors: Selectors for retrieving feature state.
• store/featureName/reducers: Pure functions that handle state changes.
• store/featureName/feature-entitys.service.ts: Service for accessing entities via NgRx Data patterns.

CompreFace offers several custom-builds optimized for different hardware (CPU vs GPU) and performance requirements (speed vs accuracy). When choosing a build, consider your hardware capabilities (AVX/AVX2 instructions or CUDA for GPU) and whether you prioritize speed or detection/recognition accuracy.

Build Selection Guide

• General Purpose: Use FaceNet (default). It supports CPUs without AVX2.
• High Speed (CPU): Use Mobilenet. It is the fastest among CPU-only models but requires AVX2 instructions.
• High Speed (GPU): Use Mobilenet-gpu. This is the fastest overall model but requires a GPU with CUDA.
• High Accuracy (CPU): Use SubCenter-ArcFace-r100. This is the most accurate CPU model but is significantly slower.
• High Accuracy (GPU): Use SubCenter-ArcFace-r100-gpu. This provides the highest accuracy and requires a GPU with CUDA.
• Experimental (Masked): Use FaceNet Masked for experimental face recognition on masked faces.

| Custom-build                  | Base library                                              | CPU                     | GPU                 | Face detection model / accuracy on [WIDER Face (Hard)] | Face recognition model / accuracy on [LFW] | Age and gender detection                                                            | Face mask detection                                                            | Comment                                        |
|-------------------------------|-----------------------------------------------------------|-------------------------|---------------------|---------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------|--------------------------------------------------|------------------------------------------------|
| FaceNet (default)            | [FaceNet](https://github.com/davidsandberg/facenet)       | x86 (AVX instructions)  | not supported       | MTCNN / 80.9%                                                                                                             | FaceNet (20180402-114759) / 99.63%                                                                                    | Custom, the model is taken [here](https://github.com/GilLevi/AgeGenderDeepLearning) | [Custom model](../docs/Mask-detection-plugin.md) | For general purposes. Support CPU without AVX2 |
| FaceNet Masked (Experimental) | [FaceNet](https://github.com/davidsandberg/facenet)       | x86 (AVX instructions)  | not supported       | MTCNN / 80.9%                                                                                                             | inception_v3_on_mafa_kaggle123 / 98.73%                                                                               | Custom, the model is taken [here](https://github.com/GilLevi/AgeGenderDeepLearning) | [Custom model](../docs/Mask-detection-plugin.md) | For general purposes. Support CPU without AVX2 |
| Mobilenet                     | [InsightFace](https://github.com/deepinsight/insightface) | x86 (AVX2 instructions) | not supported       | RetinaFace-MobileNet0.25 / 82.5%                                                                                          | MobileFaceNet,ArcFace / 99.50%                                                                                          | InsightFace                                                                         | [Custom model](../docs/Mask-detection-plugin.md) | The fastest model among CPU only models        |
| Mobilenet-gpu                 | [InsightFace](https://github.com/deepinsight/insightface) | x86 (AVX2 instructions) | GPU (CUDA required) | RetinaFace-MobileNet0.25 / 82.5%                                                                                          | MobileFaceNet,ArcFace / 99.50%                                                                                         | InsightFace                                                                         | [Custom model](../docs/Mask-detection-plugin.md) | The fastest model                              |
| SubCenter-ArcFace-r100        | [InsightFace](https://github.com/deepinsight/insightface) | x86 (AVX2 instructions) | not supported       | retinaface_r50_v1 / 91.4%                                                                                                 | arcface-r100-msfdrop75 / 99.80%                                                                                       | InsightFace                                                                         | [Custom model](../docs/Mask-detection-plugin.md) | The most accurate model, but the most slow     |
| SubCenter-ArcFace-r100-gpu    | [InsightFace](https://github.com/deepinsight/insightface) | x86 (AVX2 instructions) | GPU (CUDA required) | retinaface_r50_v1 / 91.4%                                                                                                 | arcface-r100-msfdrop75 / 99.80%                                                                                       | InsightFace                                                                         | [Custom model](../docs/Mask-detection-plugin.md) | The most accurate model                        |

Application roles define a user's permissions within a specific application. To be a member of an application team, a user must first have the Global User role and then be added to the application by a Global Owner, Global Administrator, or Application Owner.

Application roles include:

• Application Owner: The user who creates the application. Has full rights to manage the application, its users, and Face Services. An application owner cannot delete themselves without first assigning the role to someone else.
• Application Administrator: (Requires Global User role) Can create and manage Face Services but cannot manage the application settings or its users.
• Application User: (Requires Global User role) The least permissive role. This user cannot manage anything within the application but has sufficient access to integrate CompreFace with other applications.