TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection
Tony Tran, Bin Hu · 2025
Read TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection by Tony Tran, Bin Hu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-20746.txt.
Introduction
Tony Tran and Bin Hu teach tiny devices to spot litter. Trash detection on the TACO dataset must run on microcontrollers and edge hardware, so the detector family has to be both accurate and frugal. Their approach searches architectures iteratively with hardware in the loop, using a once-for-all style supernet built from residual detectors.
Evolutionary search alternates between optimizing the backbone and the neck and head, steadied by a population passthrough mechanism and an accuracy predictor that trims search cost. The resulting TrashDet family spans from about one million to thirty million parameters, trading accuracy for budget smoothly. The largest variant tops prior detectors on the five-class task with far fewer parameters. On a real microcontroller dataset, specialized variants slash energy and latency against the benchmark.
Catalog record
| CREATOR | Tony Tran, Bin Hu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-20746
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
TrashDet: Iterative Neural Architecture Search for Efficient Waste Detection Tony Tran, Bin Hu 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 7,525 μ \mu 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. 1 Introduction Uncontrolled waste disposal and littering contribute directly to environmental pollution and degradation of public spaces. Automatic image based waste detection and classification enables large scale monitoring in settings such as streets, parks, rivers, and coastlines [ 20 ] . In practice, these systems must operate continuously on battery powered or solar powered nodes, where network connectivity is unreliable and maintenance is costly. This makes real time detection on devices with very limited memory, compute, and energy a central requirement rather than an implementation detail [ 19 , 25 ] . Datasets such as TACO (Trash Annotations in Context) [ 21 ] provide realistic benchmarks for this problem, with cluttered backgrounds, small and deformable objects, and strong class imbalance across waste categories. Recent methods designed for solid waste and aerial litter, including SWDet and AltiDet [ 30 , 12 ] , improve detection quality on TACO and related datasets, but rely on backbones with tens of millions of parameters and GPU class hardware. Directly deploying such models on TinyML platforms such as microcontrollers is infeasible due to strict limits on convolutional operators, activation memory, and total layer count. Figure 1 : Comparisons with other state-of-the-art methods on the TACO [ 21 ] 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. To address this gap, we employ an iterative hardware-aware NAS framework for real-time waste detection on constrained edge devices [ 26 ] . It combines a Once-For-All (OFA) style weight-sharing detection supernet with an evolutionary search that alternates between backbone and head optimization under explicit resource limits. Population passthrough and a lightweight accuracy predictor maintain stability and efficiency in the non-convex combinatorial design space. This process yields a family of deployment-ready detectors, termed TrashDet , tailored to TACO and TinyML hardware budgets. The methodology consists of the following components: • Hardware-Constrained Supernet: We utilize a hardware-constrained OFA detection supernet for trash recognition on the TACO dataset. The design supports flexible depth, width, and expansion-ratio configurations while strictly enforcing edge-device requirements, including supported operators, flash and activation-memory limits, and other deployment constraints. • Modular Search Strategy: We adopt the iterative evolutionary NAS procedure [ 26 ] , alternating backbone and head optimization under separate computational budgets. This modular search formulation reduces the complexity of the overall design space while still enabling coupled backbone–head architectures suitable for microcontroller deployment. • Search Stabilization: To improve convergence and reduce search cost, we employ population passthrough and a learned accuracy predictor. These components stabilize the alternating search process and significantly reduce dependence on repeated full mAP50 evaluations. • Evaluation and Deployment: We evaluate the resulting TrashDet models on a five-class subset of TACO (paper, plastic, bottle, can, cigarette). The largest 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, as shown in Figure 1 . The full TrashDet family spans 1.2M to 30.5M parameters with mAP50 values between 11.4 and 19.5, providing scalable options for diverse TinyML deployment budgets. On the MAX78002 microcontroller with the TrashNet dataset, two specialized variants, TrashDet–ResNet and TrashDet–MBNet , jointly outperform the ai87-fpndetector baseline: TrashDet–ResNet achieves 7,525 \mu J per inference at 26.7 ms latency and 37.45 FPS with 1.08M parameters, while TrashDet–MBNet attains 93.3 mAP50 at 17,581 \mu J and 51.1 ms, yielding energy, latency, and power reductions of up to 88%, 78%, and 53%, respectively, compared to the baseline. 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. 2 Related Work 2.1 Object Detection Modern object detection has progressed from two stage region based frameworks to one stage and transformer based detectors. Early methods such as R-CNN and Faster R-CNN [ 5 , 24 ] used deep convolutional backbones with region proposals. Single stage detectors such as YOLO and SSD [ 22 , 16 ] improved speed by predicting bounding boxes and class scores directly from dense feature maps. Later models including YOLOv3 [ 23 ] , YOLOv5 [ 11 ] , YOLOv8 [ 10 ] , and RTMDet [ 17 ] refine the architecture and training setup to improve the accuracy and speed trade off. Transformer based methods such as DETR and Deformable DETR [ 3 , 31 ] further simplify the detection pipeline by casting detection as a set prediction problem and using attention based feature aggregation. 2.2 General Waste Detection Waste management, illegal dumping, and environmental pollution create a need for automatic systems that can detect and classify waste in real scenes. Recent work uses deep learning and computer vision to support waste sorting, litter detection, and contamination monitoring, but still reports challenges in data quality, domain shift, and deployment cost [ 20 ] . Public datasets such as TACO (Trash Annotations in Context) [ 21 ] offer in the wild images of litter across many outdoor settings and waste categories and are now standard benchmarks for trash detection and segmentation. Several methods target waste detection in aerial and outdoor imagery. SWDet [ 30 ] introduces an anchor based detector for solid waste in aerial images, with an Asymmetric Deep Aggregation backbone that improves multi scale feature representation for small and sparse trash instances. AltiDet [ 12 ] adds altitude informed feature fusion and a high resolution feature extraction backbone to handle viewpoint and scale changes in UAV imagery. Many applications such as smart bins, UAV based monitoring, and low cost