SeetaPsych Face Hub

Community open-source face modules for SeetaPsych

简体中文 | English

Usage

This project is already included in the seetapsych-lib default configuration. Download and use it via seetapsych-manager download.

For usage, refer to SeetaPsych.

You can additionally add this algorithm module using the following methods.

WebUI

Run seetapsych-webui with the --dirs argument to use it.

seetapsych-webui --files seetapsych_face_hub/modules/insightface/retinaface.yml

Programmatic Usage

Add the following code in your program to use this algorithm module.

from seetapsych_lib.runtime.factory import Factory
from seetapsych_lib.runtime.pipeline import Pipeline

factory = Factory()
factory.load_file_modules("seetapsych_face_hub/modules/insightface/retinaface.yml")

pipeline = Pipeline(factory, ...)

pipeline.add_attributes("face/detection")

For complete end-to-end examples with visualization, see:

• examples/image_retinaface.py — static-image face detection + 5-point landmarks via InsightFace RetinaFace.
• examples/image_mediapipe.py — static-image face detection + 468-point 3D mesh via MediaPipe Tasks.
• examples/image_arcface.py — static-image RetinaFace + 512-dim ArcFace embedding with voiceprint-style spectrum visualization.
• examples/camera_selection.py — live-camera face detection + PID-based target selection & tracking.

Module Catalog

InsightFace (RetinaFace)

InsightFace Buffalo RetinaFace detector: face bounding box + 5-point landmark detection with 500M / 2.5G / 10G FLOPs model variants.

Figure 1. InsightFace RetinaFace output visualization — detected faces, confidence scores, and 5-point landmarks.

Module config: insightface/retinaface.yml

Description: InsightFace RetinaFace detecting face boxes + 5-point landmarks with configurable input size; 3 accuracy/speed model tiers

Parameters

Models

Output Attributes

• face/detection — spec.
• face/landmarks — spec.

InsightFace (ArcFace)

InsightFace Buffalo face feature extraction via ArcFace models producing L2-normalized 512-dim embeddings.

Figure 2. InsightFace ArcFace output visualization — detected face and vertical voiceprint-style 512-dim embedding spectrum with sign-magnitude heat bars.

Module config: insightface/arcface.yml

Description: Extract L2-normalized 512-dim face embeddings for recognition, clustering, or similarity search from 5-point aligned faces

Parameters: (none)

Models

Output Attributes

• face/feature — spec.

MediaPipe

MediaPipe Tasks face detection and 468-point 3D face mesh landmarking via BlazeFace short/full-range models.

Figure 3. MediaPipe Tasks output visualization — face detection bbox and 468-point 3D FaceMesh landmarks.

Module config: mediapipe.yml

Package: FaceDetection-MediaPipe

Description: Face bounding box detection via MediaPipe BlazeFace detector with short/full-range/sparse model variants

Parameters

Models

Output Attributes

• face/detection — spec.

Package: FaceMesh-MediaPipe

Description: 468-point normalized 3D face mesh landmark extraction from detected face boxes with optional blendshapes and pose matrices

Parameters

Models

Output Attributes

• face/mesh — spec.

References

• MediaPipe
• InsightFace
• Pytorch_Retinaface