TL;DR
This paper introduces a method to optimize object detection models for edge devices, specifically targeting an AI assistant for the visually impaired called OLIV. By treating neural network micro-architecture design as a numerical optimization problem, the authors find an optimal balance between accuracy, speed, and model size. The resulting MobileNetV2-SSD model runs locally on-device, preserving user privacy and reducing costs without sacrificing critical performance.
The Challenge: Edge-Based Assistive Vision
Visually impaired individuals can greatly benefit from computer vision tools that help them locate displaced objects in their environment, such as finding a misplaced mouse on a desk. However, state-of-the-art object detection models are typically massive and computationally hungry, requiring powerful GPUs.
This creates a dilemma for assistive devices: running these models in the cloud raises significant privacy concerns (streaming video of a user's home) and ongoing server costs, while running them locally on affordable edge devices is difficult due to hardware constraints. The authors tackle the challenge of designing a compact, fast, and accurate object detection network that can run entirely on-device for the OLIV system (an AI-powered assistant for object localization).
Methods: Micro-architecture Design Exploration
The authors employ a "micro-architecture design exploration" strategy to optimize a MobileNetV2-SSD object detection network. Instead of manually guessing the best network size, they formulate the architecture search as a numerical optimization problem. They specifically tune two key hyperparameters:
- Width multiplier ($\alpha$): Scales the width of the network uniformly by multiplying the number of input and output channels at each layer.
- Resolution multiplier ($\rho$): Scales the input image resolution.
To evaluate the trade-offs of different $\alpha$ and $\rho$ combinations, they use a modified NetScore objective function. NetScore quantitatively balances a network's accuracy, architectural complexity (number of parameters), and computational complexity (CPU running time). Because accuracy is paramount for a reliable assistive device, the authors heavily weight the accuracy metric compared to model size and speed in their objective function.
Datasets
The research utilizes MSCOCO-OLIV, a custom 21-class subset of the standard COCO2017 dataset. This subset specifically isolates indoor objects relevant to the daily lives of visually impaired users, rather than the broader, general-purpose categories found in the full COCO dataset.
Key Insights and Results
The optimization process identified the absolute optimal MobileNetV2-SSD configuration for the OLIV system: a width multiplier of 1.3 and an input resolution of 224.
Key findings from the exploration include:
- Diminishing Returns in Scaling: Increasing the width multiplier and input resolution generally improves Mean Average Precision (mAP) and increases CPU run-time. However, the authors found a clear inflection point: after a width multiplier of 1.15 and a resolution of 220, accuracy gains severely plateau, while computational costs continue to rise linearly.
- Objective-Driven Search: By mathematically defining the ideal trade-off via the modified NetScore, the researchers could systematically pinpoint the exact hyperparameter pairing that maximizes utility for edge-based assistive tech. This configuration achieved the highest modified NetScore of 68.7.
- Privacy through Efficiency: By successfully shrinking the model to run on a local CPU without losing critical accuracy, the system inherently protects user privacy—a major hurdle for the adoption of always-on cameras in domestic environments.
Limitations and Future Work
The paper is a brief exploration and does not include a dedicated limitations section. For instance, it does not detail the specific hardware used for the CPU benchmarking or how the system performs in real-world, poorly lit environments. However, the authors note that this micro-architecture exploration approach has the potential to be applied to other assistive devices beyond OLIV, suggesting a future direction of generalizing this optimization pipeline to different hardware constraints and novel use cases.
Summary updated