solution
Tiny AI Edge Model Suite
YOLOv8n + MobileNetV3-Small with INT8 quantisation and hardware-aware NAS for Mission-specific detection.
Hardware-aware compression pipeline targeting <20 MB post-INT8 models, <100 ms inference, and >85% mAP on Mission-held-out sets (COCO/VisDrone/FLIR base + site fine-tunes).
Stage 1: baseline selection (YOLOv8n detection; MobileNetV3-Small classification).
Stage 2: INT8 PTQ (~4× size, 2–4× latency; ~0.5–2% mAP cost).
Stage 3: MCUNet-style NAS for soil/ocean models with structured features.
Mission-specific models: climate/smart-city thermal & air quality; ocean/water surface; soil health indicators; environmental carcinogen proxy.
Source: SWARMind Drones Proposal Draft § Tiny AI Model Development.
Stage 2: INT8 PTQ (~4× size, 2–4× latency; ~0.5–2% mAP cost).
Stage 3: MCUNet-style NAS for soil/ocean models with structured features.
Mission-specific models: climate/smart-city thermal & air quality; ocean/water surface; soil health indicators; environmental carcinogen proxy.
Source: SWARMind Drones Proposal Draft § Tiny AI Model Development.
Mission profile
Fine-tune with ≥200 site images and ~4 GPU-hours per the replication framework module.