Robotic Object Classification: Deep Learning vs. Local Features

This coursework compares deep learning approaches with local feature-based pipelines for robotic object classification. The report evaluates robustness under realistic conditions such as viewpoint changes, lighting variation, and partial occlusion.

It discusses practical trade-offs between classical and learned representations, including dataset demands, generalisation behaviour, and deployment considerations for autonomous robotic systems. I also used Grad-CAM analysis to inspect model attention and diagnose where deep models were relying on robust object features versus background cues.