Memento-Face 0.1.7

Easy to use Face Recognition and Detection models.

---

Face Detection and Recognition

A PyTorch-based face detection and recognition system built with EfficientNet backbones and ArcFace loss, supporting live webcam detection and recognition, image inference, and easy embedding extraction.

---

Table of Contents

• Overview

• Model Architecture

• Installation

• Usage

  • Detection Model
  • Recognition Model
  • Live Webcam Testing

• Saving and Loading Weights

• Customizing Confidence Thresholds and Frame Skips

• Contributing

• License

---

Overview

This repository implements a face detection and recognition pipeline consisting of:

• DetectionModel: Uses EfficientNet-B0 as backbone, with a custom neck and heads for bounding box and objectness prediction.
• RecognitionModel: Uses EfficientNet-B2 backbone, neck, pooling, and ArcFace for robust face embeddings and classification.

---

Model Architecture

DetectionModel

• Backbone: EfficientNet-B0 feature extractor (pretrained weights).

• Neck: Three Conv-BatchNorm-SiLU blocks reducing channel depth from 1280 to 96.

• Heads:

  • BBox Head: Predicts bounding box coordinates (4 channels).
  • Objectness Head: Predicts objectness confidence (1 channel).

• Output: Concatenated tensor of shape [B, 5, H, W] (bbox + objectness).

• Decoding: Converts network outputs to bounding boxes and confidence scores using grid and stride calculations.

RecognitionModel

• Backbone: EfficientNet-B2 feature extractor.
• Neck: Conv-BatchNorm-SiLU block to reduce channels to embedding dimension (default 256).
• Pooling: Adaptive average pooling to 1x1.
• Embedding Head: BatchNorm and Dropout followed by normalization.
• ArcFace: Angular margin softmax layer for face recognition classification.
• Output: Either normalized embeddings or classification logits (if labels provided).

---

Installation

To install from PyPI:

pip install MementoML

Make sure you have Python 3.8+ and PyTorch installed. Install dependencies:

I used Python 3.12.7.

pip install torch torchvision numpy opencv-python matplotlib pillow

or

pip install -r requirements.txt

Place the pre-trained weight files in the working directory:

• FaceDetectionWeights.pth
• FaceRecognitionWeights.pth

---

Usage

Detection Model

Detect faces in a single image and plot bounding boxes:

from PIL import Image
import matplotlib.pyplot as plt

detector = DetectionModel(weights="FaceDetectionWeights.pth", device="cuda")
img = Image.open("test_face.jpg")
detector.face_and_plot(img, conf_thresh=0.5)

Run live webcam face detection:

detector = DetectionModel(weights="FaceDetectionWeights.pth", device="cuda")
detector.live_test(conf_thresh=0.8, frame_skip=0)

---

Recognition Model

Generate a face embedding from an image:

from PIL import Image

recognizer = RecognitionModel(weights="FaceRecognitionWeights.pth", device="cuda")
img = Image.open("face_crop.jpg")
embedding = recognizer.generate_emb(img)
print(embedding.shape)  # torch.Size([256])

---

Live Webcam Testing

Both detection and recognition models support live webcam testing individually:

Detection example shown above; for recognition, run your own scripts on cropped face images or saved crops.

---

Saving and Loading Weights

Save your model weights after training:

detector.save_weights("new_detection_weights.pth")
recognizer.save_weights("new_recognition_weights.pth")

Load weights:

detector = DetectionModel(weights="new_detection_weights.pth")
recognizer = RecognitionModel(weights="new_recognition_weights.pth")

---

Customizing Confidence Thresholds and Frame Skips

• Confidence Threshold: Adjust detection sensitivity.

detector.live_test(conf_thresh=0.5)  # More sensitive, detect more faces

• Frame Skip: Process every Nth frame in live webcam feed to reduce compute.

detector.live_test(frame_skip=5)  # Process every 5th frame

---

Contributing

Feel free to open issues or submit pull requests. Suggestions and improvements are welcome!

---

License

This project is licensed under the MIT License.

---

If you want me to generate examples for training, or detailed info about the ArcFace loss or anything else, just ask! My email: [e-mail]

---