TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection

TL;DR

TrashDet introduces a hardware-aware neural architecture search framework that builds an OFA-style detection supernet and uses iterative evolutionary search to optimize backbone and neck/head under TinyML constraints. The approach yields a family of deployment-ready detectors spanning 1.2M to 30.5M parameters, with TrashDet-l achieving leading $mAP50$ on a five-class TACO subset while remaining compact. Specialized TinyML deployments on MAX78002, TrashDet–ResNet and TrashDet–MBNet, demonstrate dramatic reductions in energy, latency, and power compared with a baseline, highlighting practical impact for edge environmental monitoring. Overall, the work enables scalable, real-time waste detection on resource-constrained hardware through coordinated architecture search and hardware-aware design choices.

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 (paper, plastic, bottle, can, 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. The TrashDet family spans 1.2M to 30.5M parameters with mAP50 values between 11.4 and 19.5, providing scalable detector options for diverse TinyML deployment budgets on resource-constrained hardware. On the MAX78002 microcontroller with the TrashNet dataset, two specialized variants, TrashDet-ResNet and TrashDet-MBNet, jointly dominate the ai87-fpndetector baseline, with TrashDet-ResNet achieving 7525~$μ$J energy per inference at 26.7 ms latency and 37.45 FPS, and TrashDet-MBNet improving mAP50 by 10.2%; together they reduce energy consumption by up to 88%, latency by up to 78%, and average power by up to 53% compared to existing TinyML detectors.

Figures (4)

• Figure 1: Comparisons with other state-of-the-art methods on the TACO taco2020 dataset in terms of size-accuracy, reducing model size by 2.8x for similarly performing models and increasing accuracy by 2.0% against similarly sized models.
• Figure 2: Overview of TrashDet framework. A unified OFA-style supernet is first constructed, comprising a ResNet-style backbone, neck, and YOLO-style detection head (left). Iterative evolutionary search is then performed in two coordinated stages: searching for the optimal backbone while keeping the neck/head fixed (Step 2), and searching for the optimal neck/head while fixing the discovered backbone (Step 3). The evolutionary controller evaluates candidate subnets using detection performance on TACO and alternates between these stages to obtain compact, deployment-ready TrashDet models.
• Figure 3: Accuracy predictor versus true mAP50 for candidate subnets. Points lying close to the dashed identity line indicate that the predictor is a reliable surrogate for detection performance during search.
• Figure 4: Basic Building Blocks. (a) The original bottleneck block used in ResNets he2016deep. (b) The DarkNet block used in CSPDarknets wang2020cspnet. (c) RepVGGBlock block used in RTMDet lyu2022rtmdet. (d) The OFA-Res Block Used in OFA ResNet cai2020once.