Abstract: This paper addresses trash detection on the TACO dataset under strict TinyML constraints using an iterative hardware-aware neural architecture search framework targeting edge and IoT devices. The proposed method constructs a Once-for-All-style ResDets supernet and performs iterative evolutionary search that alternates between backbone and neck/head optimization, supported by a population passthrough mechanism and an accuracy predictor to reduce search cost and improve stability. This framework yields a family of deployment-ready detectors, termed TrashDets. On a five-class TACO subset consisting of paper, plastic, bottle, can, and cigarette, the strongest variant, TrashDet-l, achieves 19.5 mAP50 with 30.5M parameters, improving accuracy by up to 3.6 mAP50 over prior detectors while using substantially fewer parameters. On the MAX78002 microcontroller with the TrashNet dataset, two specialized variants, TrashDet-ResNet and TrashDet-MBNet, jointly dominate the ai87-fpndetector baseline. TrashDet-ResNet reaches 7,525 μJ energy per inference at 26.7 ms latency and 37.45 FPS, while TrashDet-MBNet improves mAP50 by 10.2 points. Together they reduce energy by up to 88%, latency by up to 78%, and average power by up to 53% compared to prior TinyML detectors.
<!DOCTYPE html>
Redirecting to the TrashDet project page...