Foundation-Model Distillation for Edge Inference

Distilling large vision encoders into compact students that run on a Raspberry Pi, evaluated on pest detection. Final-year project, in progress.

Status: in progress, collaborative. My final-year project, with Prof. Faisal Shafait (NUST) and Dr. Christian Weis (RPTU Kaiserslautern-Landau). I began it during a research internship at RPTU (8 Jun – 5 Sep 2026).

Question

Vision foundation models produce strong general features, but they are far too large for battery-powered field devices. Which teacher is worth distilling, and what does a small student inherit from it, including its shortcuts? We use agricultural pest detection and classification as a realistic evaluation task with tight hardware budgets.

Done so far

  • Established YOLO detection baselines for a five-class pest detection task, as part of a joint benchmarking effort.
  • Evaluated frozen teacher encoders with linear classification heads to compare how well candidate foundation-model teachers transfer to the task.
  • Ran on-device inference on a Raspberry Pi 5 and measured latency and power consumption, so that model choices are judged against real energy and hardware constraints, not just FLOPs.

Next

  • Design the student architecture and distillation objective.
  • Study which teacher properties (e.g., reliance on background or scene context) carry over to the student.