Abstract
Object detection is an important problem in a wide variety of computer vision applications for sustainable smart cities. Deep neural networks (DNNs) have attracted increasing interest in object detection due to their potential to provide high accuracy detection performance in challenging scenarios. However, DNNs involve high computational complexity and are therefore challenging to deploy under the tighter resource constraints of edge cloud environments compared to more resource-abundant platforms, such as conventional cloud computing platforms. Moreover, the monolithic structure of conventional DNN implementations limits their utility under the dynamically changing operational conditions that are typical in edge cloud computing. In this paper, we address these challenges and limitations of conventional DNN implementation techniques by introducing a new resource-adaptive scheme for DNN-based object detection. This scheme applies the recently-introduced concept of elastic neural networks, which involves the incorporation of multiple outputs within intermediate stages of the neural network backbone. We demonstrate a novel elastic DNN design for object detection, and we show how other methods for streamlining resource requirements, in particular network pruning, can be applied in conjunction with the proposed elastic network approach. Through extensive experiments, we demonstrate the ability of our methods to efficiently trade-off computational complexity and object detection accuracy for scalable deployment.
Introduction
Modern object detection relies on Deep Neural Networks (DNNs) which are notoriously "monolithic." Once a model is trained and deployed on an edge device—such as a drone or a traffic camera—its computational cost is fixed. This rigidity poses a significant challenge in edge cloud environments where operational conditions change dynamically. For instance, if a drone’s battery is low or if it needs to process frames faster to track a speeding vehicle, a standard DNN cannot easily "scale down" its complexity to conserve power or reduce latency.
Current workarounds, such as maintaining multiple models of varying sizes on a single device, are inefficient due to the high storage and memory overhead. This paper introduces ESNO (Elasticized SSD Network for Object detection), a resource-adaptive framework that allows a single model to change its complexity on the fly, effectively trading off accuracy for computational efficiency.
Methodology
The ESNO framework is built upon the popular Single-Shot Multibox Detector (SSD) using a VGG-16 backbone. The system employs two primary strategies to achieve resource adaptability:
- Elastic Neural Networks (ENNs): The architecture incorporates "early exits" at intermediate stages of the network backbone. Depending on the available resources, an image can either pass through the full depth of the network for maximum precision or exit early through a "post-exit subnetwork" to produce a faster, albeit slightly less accurate, result.
- Network Pruning: To mitigate the increased parameter count caused by adding multiple exit heads, the authors applied L1Unstructured pruning. This technique identifies and removes the least significant weights, allowing the entire multi-exit system to fit within the memory footprint of a standard, non-elastic model.
The training process involves a two-step approach: the full network is trained first, followed by "freezing" the backbone while training the individual early-exit heads to interpret intermediate features. The models were evaluated using the Pascal Visual Object Classes (VOC) 2007 and 2012 datasets, which are standard benchmarks for object localization and classification.
Results and Discussion
The researchers tested five distinct execution paths, labeled ESSD1 (the shallowest/fastest) through ESSD5 (the full-depth model).
Performance Trade-offs
The experiments revealed a clear and manageable trade-off between speed and accuracy. The earliest exit, ESSD1, utilized approximately 31% fewer parameters than the full model. While the full unpruned model (ESSD5) achieved a Mean Average Precision (mAP) of 77.49, the ESSD1 path still maintained a respectable 70.00 mAP.
Pruning Synergy
A key finding was the high complementarity between pruning and elasticity. With 30% pruning, the entire elastic system occupied roughly the same memory as the original SSD while maintaining accuracy within 2% mAP of the unpruned versions. Interestingly, in the case of ESSD4, pruning slightly improved accuracy, likely due to a regularizing effect on the network.
Graceful Degradation
The ESNO framework enables "graceful degradation," where the system remains functional under resource constraints by switching to a lower-complexity path. This is particularly valuable for smart city applications where "good enough" detection is often preferable to total system failure during power-saving modes.
Conclusion
This work demonstrates that elastic neural networks, combined with strategic pruning, provide a viable path for deploying sophisticated AI on resource-constrained edge devices. While the current study utilizes manual exit placement, future research will focus on automated optimization of exit locations and the development of dynamic runtime controllers to manage switching logic based on real-time environmental factors.
Summary updated