[Paper Review] YOLO Nano: a Highly Compact You Only Look Once Convolutional Neural Network for Object Detection

Alexander Wong, Mahmoud Famuori+4 more|arXiv (Cornell University)|Oct 2, 2019

Advanced Neural Network Applications4 citations

TL;DR

YOLO Nano is a highly compact, human-machine collaborative object detection network designed for edge and mobile deployment, achieving 69.1% mAP on VOC 2007 with only 4.0MB model size and 4.57B FLOPs—outperforming Tiny YOLOv2 and Tiny YOLOv3 in accuracy while reducing size and computation by over 8× and 34%, respectively.

ABSTRACT

Object detection remains an active area of research in the field of computer vision, and considerable advances and successes has been achieved in this area through the design of deep convolutional neural networks for tackling object detection. Despite these successes, one of the biggest challenges to widespread deployment of such object detection networks on edge and mobile scenarios is the high computational and memory requirements. As such, there has been growing research interest in the design of efficient deep neural network architectures catered for edge and mobile usage. In this study, we introduce YOLO Nano, a highly compact deep convolutional neural network for the task of object detection. A human-machine collaborative design strategy is leveraged to create YOLO Nano, where principled network design prototyping, based on design principles from the YOLO family of single-shot object detection network architectures, is coupled with machine-driven design exploration to create a compact network with highly customized module-level macroarchitecture and microarchitecture designs tailored for the task of embedded object detection. The proposed YOLO Nano possesses a model size of ~4.0MB (>15.1x and >8.3x smaller than Tiny YOLOv2 and Tiny YOLOv3, respectively) and requires 4.57B operations for inference (>34% and ~17% lower than Tiny YOLOv2 and Tiny YOLOv3, respectively) while still achieving an mAP of ~69.1% on the VOC 2007 dataset (~12% and ~10.7% higher than Tiny YOLOv2 and Tiny YOLOv3, respectively). Experiments on inference speed and power efficiency on a Jetson AGX Xavier embedded module at different power budgets further demonstrate the efficacy of YOLO Nano for embedded scenarios.

Motivation & Objective

• Address the challenge of deploying large, computationally expensive object detection networks on memory- and power-constrained edge and mobile devices.
• Overcome the limitations of existing compact models like Tiny YOLOv2 and Tiny YOLOv3, which sacrifice accuracy for size and speed.
• Develop a highly efficient, compact object detection architecture through a human-machine collaborative design strategy that combines principled prototyping with machine-driven architecture search.
• Optimize for real-world embedded deployment by balancing model size, inference speed, and accuracy under constrained power budgets.
• Demonstrate superior performance on embedded hardware, particularly in power efficiency and inference speed on Jetson AGX Xavier.

Proposed method

• Employ a two-stage human-machine collaborative design strategy: (1) principled network design prototyping based on YOLO family design principles, and (2) machine-driven design exploration using generative synthesis.
• Leverage a custom search space that enables fine-grained architectural customization at both macro- and micro-architecture levels, including unique module compositions and layer-specific designs.
• Introduce lightweight fully-connected attention (FCA) modules strategically within the network via machine-driven exploration, enabling dynamic feature recalibration without significant computational overhead.
• Use a differentiable, differentiable architecture search (DARTS)-inspired approach to explore diverse architectural configurations, including PEP modules, EP modules, FCA, and standard 1×1 and 3×3 convolutions.
• Optimize for edge deployment by co-designing model size, FLOPs, and inference speed through power-aware evaluation on Jetson AGX Xavier.
• Apply channel-wise feature recalibration via FCA, which uses two fully-connected layers to learn dynamic inter-dependencies between channels and reweight features for improved expressiveness.

Experimental results

Research questions

• RQ1 Can a human-machine collaborative design strategy produce a more compact and accurate object detection network than existing state-of-the-art tiny models?
• RQ2 To what extent can architectural heterogeneity at both macro- and micro-levels improve the balance between model expressiveness, computational cost, and size?
• RQ3 How does the integration of lightweight attention mechanisms (e.g., FCA) impact performance and efficiency in ultra-compact networks?
• RQ4 Can a highly compact network achieve superior accuracy and efficiency compared to Tiny YOLOv2 and Tiny YOLOv3 on edge hardware under constrained power budgets?
• RQ5 What is the real-world inference speed and power efficiency of the proposed network on embedded platforms like Jetson AGX Xavier?

Key findings

• YOLO Nano achieves a model size of 4.0MB, which is over 15× smaller than Tiny YOLOv2 (60.5MB) and over 8× smaller than Tiny YOLOv3 (33.4MB).
• The network requires only 4.57 billion FLOPs for inference, representing a 34% reduction compared to Tiny YOLOv2 (6.97B FLOPs) and a 17% reduction compared to Tiny YOLOv3 (5.52B FLOPs).
• YOLO Nano achieves 69.1% mAP on the VOC 2007 test set, which is 12% higher than Tiny YOLOv2 (57.1%) and 10.7% higher than Tiny YOLOv3 (58.4%).
• On the Jetson AGX Xavier, YOLO Nano achieves 26.9 FPS at 15W and 48.2 FPS at 30W, with power efficiencies of 1.97 and 1.61 images/sec/watt, respectively.
• The machine-driven design process successfully generated a highly heterogeneous architecture with unique macro- and micro-architectures across modules, enabling optimal trade-offs between complexity and expressiveness.
• The strategic integration of lightweight fully-connected attention (FCA) modules via automated search enhanced feature recalibration and contributed to improved accuracy without increasing computational cost.