Face-detection-and-recognition-with-YOLO

This project implements a custom YOLO (You Only Look Once)👀 model for detecting faces in images. The model has been built from scratch using PyTorch, and it predicts bounding boxes around faces in images.

1. Model Architecture:

   • The YOLO architecture consists of convolutional layers for feature extraction and fully connected layers for predicting bounding boxes.
   • The model outputs grid cells (14x14), each predicting 1 bounding box along with confidence score.

2. Loss Function:

   • The custom loss function includes multiple components:
     • Coordinate Loss (lambda_coord=5): Penalizes errors in the predicted center coordinates of the bounding boxes.
     • Size Loss (lambda_size=5): Penalizes discrepancies in the predicted width and height of the bounding boxes.
     • Object Loss (lambda_obj=1): Penalizes differences in confidence scores for boxes that contain objects.
     • No Object Loss (lambda_noobj=0.5): Reduces false positives by penalizing confidence in grid cells without objects.

3. Non-Maximum Suppression (NMS):

   • After the model generates bounding box predictions, NMS is applied to eliminate redundant boxes.
   • NMS filters out overlapping boxes based on Intersection over Union (IoU) and retains only the most confident predictions.

Directory Structure

• 00_FaceDetection: Contains the final implementation with all scripts for training, evaluation, and inference.
• previous_attempts: Holds earlier versions of the project, documenting issues with dataset construction and loss function design.

Development Process

Constructed models:

• https://drive.google.com/drive/folders/1YU3ZbLDScdlr_SaJtzAjiH6ksqnjWmME?usp=drive_link

Data:

• https://drive.google.com/drive/folders/1WhAjMJtyc5AcPpBlUCGp75LF88TnC4Q7?usp=drive_link

References:

• Redmon_You_Only_Look_CVPR_2016_paper.pdf
• face_detection_and_recognition.pdf

Student:

• Vasilije Todorović