On-Device Real-Time Blurring
Real-time privacy blurring at target FPS on-device under thermal and power constraints — 1st Prize, SNU Ambient AI Competition.
1st Prize — SNU Ambient AI Team of 5 Aug–Sep 2024
Problem
Privacy-sensitive blurring (faces, license plates) must run on-device — no cloud round-trips. The hard constraint: hitting a target FPS on a mid-range Android phone while the battery drains and the SoC throttles.
Most approaches optimize for accuracy in isolation. We had to co-optimize model size, runtime backend, and degradation policy simultaneously.
Approach
Runtime
- Evaluated PyTorch Mobile vs NNAPI backend across thermal states
- Implemented graceful degradation — reduced resolution at high thermal load to sustain FPS
Model optimization
- Explored quantization strategies (INT8, FP16) and input scaling trade-offs
- Built a deterministic latency + quality harness for controlled comparison
Results
| Config | FPS (normal) | FPS (throttled) | Quality |
|---|---|---|---|
| Baseline (FP32) | target | drops | high |
| INT8 quantized | ✅ target | ✅ target | acceptable |
| + degradation policy | ✅ target | ✅ sustained | adaptive |
Delivered a demo app and benchmark report. Coordinated split workstreams across model, runtime, and UX subteams.