ELASTIC: Efficient Once For All Iterative Search for Object Detection on Microcontrollers

TL;DR

ELASTIC tackles the challenge of deploying accurate object detectors on TinyML devices by introducing a constrained, once-for-all NAS that alternates optimization between backbone and head to preserve cross-module interactions. The key innovations are Iterative Evolutionary Architecture Search and Population Passthrough, which maintain continuity of high-quality architectures across module switches under explicit resource budgets. Empirical results show ELASTIC achieves higher mAP and faster convergence than progressive NAS on SVHN and PascalVOC subsets, and scales to full PascalVOC with strong gains over MCUNET and TinyissimoYOLO, while lowering MACs and enabling end-to-end MCU deployment. Deployment experiments on MAX78000/MAX78002 demonstrate substantial energy and latency improvements, highlighting ELASTIC’s practical impact for real-world TinyML object detection on resource-constrained devices.

Abstract

Deploying high-performance object detectors on TinyML platforms poses significant challenges due to tight hardware constraints and the modular complexity of modern detection pipelines. Neural Architecture Search (NAS) offers a path toward automation, but existing methods either restrict optimization to individual modules, sacrificing cross-module synergy, or require global searches that are computationally intractable. We propose ELASTIC (Efficient Once for AlL IterAtive Search for ObjecT DetectIon on MiCrocontrollers), a unified, hardware-aware NAS framework that alternates optimization across modules (e.g., backbone, neck, and head) in a cyclic fashion. ELASTIC introduces a novel Population Passthrough mechanism in evolutionary search that retains high-quality candidates between search stages, yielding faster convergence, up to an 8% final mAP gain, and eliminates search instability observed without population passthrough. In a controlled comparison, empirical results show ELASTIC achieves +4.75% higher mAP and 2x faster convergence than progressive NAS strategies on SVHN, and delivers a +9.09% mAP improvement on PascalVOC given the same search budget. ELASTIC achieves 72.3% mAP on PascalVOC, outperforming MCUNET by 20.9% and TinyissimoYOLO by 16.3%. When deployed on MAX78000/MAX78002 microcontrollers, ELASTICderived models outperform Analog Devices' TinySSD baselines, reducing energy by up to 71.6%, lowering latency by up to 2.4x, and improving mAP by up to 6.99 percentage points across multiple datasets.

Figures (7)

• Figure 1: Comparison of Neural Architecture Search (NAS) strategies on the SVHN (left) and PascalVOC (right) subset. On the SVHN dataset, ELASTIC achieves 4.7% higher mAP and 2$\times$ faster than Progressive. On PascalVOC, ELASTIC achieves a 9% absolute mAP gain over all other approaches in 17.5 GPU hours.
• Figure 2: Overview of ELASTIC: Our method begins with a pretrained supernet and performs iterative neural architecture search by alternating optimization between the backbone and head. The Population Passthrough mechanism ensures continuity by retaining top-performing candidates across module alternations.
• Figure 3: Effect of Population Passthrough on Search Stability. Eliminating passthrough (0%) leads to unstable search dynamics, with mAP repeatedly dropping and stalling at 22.1%. Introducing a moderate passthrough rate (50--80%) stabilizes training and improves convergence, achieving over 30% mAP (+8%).
• Figure 4: Search space refinement through ELASTIC iteration. Distributions of 100 randomly sampled architectures from three head search spaces—joint NAS, progressive head-only, and ELASTIC-refined—on SVHN (left) and PascalVOC (right). ELASTIC produces architectures with a mean mAP improvement of +4.87% on SVHN and +0.43% on PascalVOC over progressive search. Compared to global joint search, ELASTIC improves the mAP by 26.07% and 23.33% on SVHN and PascalVOC, respectively.
• Figure 5: Search Space Evolution Over Iterations. Mean accuracy and distribution of randomly sampled architectures at iterations 0, 3, and 5. On SVHN, ELASTIC continues to improve mean mAP by +3.3% while reducing variance by approximately 89.7%. On PascalVOC, mean mAP improves from 27.89% to 28.32%, with variance dropping by approximately 33%.