Model Architecture

• Implemented ResNet34 as the primary backbone for feature extraction.
• Explored ResNet18, ResNet50, and YOLOv8-pose architectures for comparative performance analysis.
• Trained models using 5-fold cross-validation, ensuring robustness and generalization.

Data Pipeline

• Full end-to-end pipeline for image preprocessing, augmentation, and dataset splitting.
• Leveraged OpenCV, NumPy, and Pandas for image transformations, normalization, and dataset management.

Training & Optimization

• Advanced training strategies including learning rate scheduling, data augmentation, and batch normalization.
• Monitored training with Matplotlib visualizations for loss curves, accuracy, and landmark heatmaps.
• Implemented custom loss functions for landmark regression tasks.

GUI Application

• Built a Tkinter-based GUI to allow non-technical users to upload images and visualize predicted landmarks.
• Integrated the trained model inference pipeline directly into the GUI for real-time predictions.