compvision_learning

Learning Computer Vision

Project 1: Human vs. Non-Human Face Classification

Description

This project focuses on developing a deep learning model to differentiate between real human faces and artificial representations. The core objective is to classify whether a detected face belongs to a Human or a Non-Human entity (encompassing anime, cartoons, video game characters, statues, and digital renders).

This task is foundational for applications in:

• Liveness Detection: Enhancing security systems against spoofing.
• Content Filtering: Automatically identifying and categorizing artistic vs. real-world content.
• Dataset Cleaning: Sanitizing large-scale face datasets for training biometric systems.

Project 2: Face Analysis in Crowds

Description

Evaluate the efficiency and accuracy of state-of-the-art face detection models (InsightFace and Uniface) in complex, high-density crowd scenarios.

Objectives

• Crowd Benchmarking: Assess how well models detect small, partially occluded, or distant faces in crowded environments.
• Comparative Analysis: Performance comparison between InsightFace (SCRFD/Buffalo_L) and Uniface (YOLOv5Face) models.
• Efficiency Metrics:
  • Detection Rate: Percentage of ground-truth faces correctly detected.
  • Count Error: Mean Absolute Error (MAE) between detected face count and ground truth.
  • Precision/Recall: Statistical evaluation of detection quality.

Dataset

• Source: Crowd Counting Dataset (Roboflow)
• Format: YOLOv5 PyTorch (Normalized bounding boxes).
• Split: Evaluated on the test set containing various crowd densities and lighting conditions.