TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection

Tony Tran 1, Bin Hu 2
1Cullen College of Engineering Research Computing at University of Houston 2Department of Engineering Technology at University of Houston
Paper Slides Video
TrashDet accuracy-size comparison

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.

Overview of the TrashDet Framework

Overview of the TrashDet framework

TrashDet first constructs a unified OFA-style detection supernet spanning the backbone, neck, and YOLO-style detection head. It then alternates evolutionary search between the backbone and neck/head under hardware constraints, using population passthrough to preserve strong candidates across iterations and produce compact, deployment-ready waste detectors.

Basic Building Blocks

Basic building blocks

TrashDet builds on OFA-style residual blocks to support dynamic specialization over depth, width, and expansion ratio while remaining compatible with deployment constraints.

Accuracy Predictor

Accuracy predictor versus true mAP50

The lightweight predictor closely tracks true mAP50 and serves as an efficient surrogate during evolutionary search, reducing the need for repeated full evaluations.

Comparison Against State-of-the-Art Detectors on TACO

Comparison of detectors on the TACO dataset. TrashDet-l achieves the highest mAP50 of 19.5, outperforming the strongest prior baseline while using roughly one third of its parameters.
Method Backbone Neck Head Params AR mAP50
YOLOv5m CSPDarknet SPPF + PANet YOLOv3 21.2M 22.3 15.9
YOLOv8m CSPDarknet SPPF + PANet YOLOv8 25.9M 16.6 16.6
TrashDet-m (Ours) OFA ResNet OFA PANet OFA YOLOv3 21.0M 19.1 18.6
SWDet-m ADA EAFPN YOLOv3 33.85M 21.0 16.4
Deformable DETR ResNet-101 DETR Encoder DETR Decoder 40M 30.3 16.8
RTMDet RTMDet-l PANet RTMDet 52.3M 19.4 16.9
AltiDet-m ADA + HRFE A-IFPN YOLOv3 85.3M 22.4 18.4
TrashDet-l (Ours) OFA ResNet OFA PANet OFA YOLOv3 30.5M 18.6 19.5

TrashDet targets a stronger accuracy-efficiency trade-off on cluttered waste scenes, where precise detection is more important than maximizing recall alone.

TrashDet Model Family on TACO

TrashDet model family illustration
Performance of TrashDet variants on TACO. The family spans deployment budgets from a 1.2M parameter ultra-compact model to a 30.5M parameter high-accuracy model.
Method Resolution Params AR mAP50 Latency (ms) FPS
TrashDet-n 640 1.2M 21.2 11.4 2.21 452.79
TrashDet-s 640 7.9M 16.9 15.8 3.83 261.06
TrashDet-m 640 21.0M 19.1 18.6 4.39 227.70
TrashDet-l 640 30.5M 18.6 19.5 5.07 197.08

All variants share the same OFA ResNet-style design space, letting practitioners choose a point on the accuracy-latency-model-size trade-off curve.

TinyML Deployment on MAX78002

TrashDet deployment results on MAX78002
Energy, latency, power, and accuracy on the MAX78002. TrashDet discovers both an efficiency-first and an accuracy-first design that dominate the baseline across multiple hardware metrics.
Model Resolution Dataset Params Energy (μJ) Latency (ms) Power (mW) FPS mAP50
ai87-fpndetector 256×320 TrashNet 2.18M 62001 122.6 445.76 8.16 83.1
TrashDet - MBNet 224×224 TrashNet 1.32M 17581 51.1 285.02 19.57 93.3
TrashDet - ResNet 224×224 TrashNet 1.08M 7525 26.7 210.5 37.45 84.6

TrashDet-ResNet provides the most aggressive efficiency setting, while TrashDet-MBNet delivers the strongest detection accuracy under the MAX78002 hardware budget.

Video Demo

BibTeX

@inproceedings{tran2026trashdet,
  title={TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection},
  author={Tony Tran and Bin Hu},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)},
  year={2026}
}