Age Estimation in the Wild

📌 Overview

Real-world age estimation suffers significantly from face occlusions, extreme lighting, and poor image quality. Traditional single-face models often degrade under these conditions.

This project solves this by shifting from localized facial analysis to holistic contextual feature extraction. By processing both facial features and body context simultaneously, the system delivers robust, real-time predictions even when the face is partially hidden.

📊 Supported Datasets

The pipeline features a unified data abstraction layer (dataset.py) that standardizes and handles on-the-fly normalization for three major benchmarks:

• IMDB-WIKI: Over 500k images for extensive pre-training.
• UTKFace: Highly diverse dataset tracking ages from 0 to 116.
• CACD: Cross-age celebrity dataset containing longitudinal variations over a decade.

🧠 Developed Methodologies

1. MiVolo (Multi-Input Vision Outlooker)

• Dual-Stream Topology: Processes standardized Face Crops and Body Crops in parallel.
• VOLO Backbone: Utilizes outlook attention for fine-grained facial textures, and class-attention layers for global contextual embedding.
• Loss Function: Optimized via Smooth L1 / MAE loss to remain robust against outlier web labels.

2. Deep ResNet Baseline

• Architecture: A standard ResNet backbone coupled with a custom linear regression head.
• Purpose: Serves as a high-speed control baseline to precisely quantify the performance gains brought by adding body context.

3. Multi-Input GhostFaceNet

• Efficiency: Designed for resource-constrained environments by deploying Ghost Convolutions (generating feature maps via cheap linear operations).
• Result: Drastically minimizes FLOPs and parameter count without losing the benefits of dual-stream context.

🛠️ Infrastructure & Features

• Experiment Tracking: Integrated with Weights & Biases (WandB) for real-time loss, metric logging, and hyperparameter tuning (hyper.py).
• Live Inference: Camera.py leverages OpenCV for dynamic bounding-box extraction and real-time webcam age tracking.
• Production-Ready Pipeline: Fully deterministic training loop with cross-epoch model checkpointing (weights, optimizer states, and schedulers).