Source-linked AI summary
YOLO Nano: a Highly Compact You Only Look Once Convolutional Neural Network for Object Detection
Alexander Wong, Mahmoud Famuori, Mohammad Javad Shafiee, Francis Li, Brendan Chwyl, Jonathan Chung
TL;DR
Object detection networks face computational and memory constraints that make edge and mobile deployment challenging. YOLO Nano combines YOLO-based prototyping with machine-driven design exploration. It is smaller than Tiny YOLOv2 and Tiny YOLOv3 while achieving higher VOC 2007 mAP, and was evaluated on an embedded module at different power budgets.
Problem
Object detection networks can be difficult to deploy on edge and mobile devices because of their computational and memory requirements.
Source: §1 · pp. 1–2Method
YOLO Nano combines principled prototyping based on YOLO single-shot design principles with machine-driven exploration of customized network designs.
Source: §1 · p. 2, §2 · p. 2, §2.1 · p. 2Results
4.0MB, >15.1× and >8.3× smaller than Tiny YOLOv2 and Tiny YOLOv3; 69.1% VOC 2007 mAP, ~12% and ~10.7% higher, respectively.
Source: §4 · p. 4Takeaways & Limitations
Jetson AGX Xavier tests at different power budgets support the paper's conclusion that YOLO Nano balances accuracy, size, and computational complexity for embedded object detection.
Source: §4 · p. 4The exploration requirements include mAP ≥65% on VOC 2007, computational cost ≤5B operations, and 8-bit weight precision.
Source: §2.2 · p. 3
Abstract
from arXiv · showhide
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.
1 Introduction
Object detection networks can be difficult to deploy on edge and mobile devices because of computational and memory constraints. YOLO Nano applies a human-machine collaborative design strategy to create a compact network for embedded object detection.
1 Introduction
- Edge and mobile deployment is constrained by the computational and memory requirements of object detection networks; Faster R-CNN can run at low single-digit frame rates on embedded processors.Local embedded processing is relevant to unmanned aerial vehicles, video surveillance, and autonomous driving.Source: §1 · p. 1
- YOLO networks use single-shot architectures, but YOLOv3 is about 240MB and requires over 65B operations, while Tiny YOLO reduces model size at a cost to detection performance.Source: §1 · p. 2
- YOLO Nano couples principled prototyping based on YOLO design principles with machine-driven exploration to create customized module-level designs for embedded object detection.Source: §1 · p. 2
2 Methods
YOLO Nano is designed through principled network prototyping followed by machine-driven design exploration. The prototype supplies a YOLO-based starting point, while exploration selects network designs within edge- and mobile-oriented requirements.
2 Methods
- The two design stages are principled network design prototyping and machine-driven design exploration.Source: §2 · p. 2
2.1 Principled network design prototyping
- The initial prototype, ϕ, uses YOLO single-shot principles, making image predictions in one forward pass rather than the hundreds to thousands of passes used by region proposal-based networks.Source: §2.1 · p. 2
- The prototype stacks feature representation modules with shortcuts and three-scale features, followed by convolutions that produce bounding-box, objectness, and class predictions at three scales.Source: §2.1 · p. 2
2.2 Machine-driven design exploration
- Generative synthesis uses the prototype, data, and edge/mobile requirements to explore module-level macroarchitecture and microarchitecture designs.Figure 1 labels PEP(x) by first-projection channels and FCA(x) by reduction ratio.Source: §2.2 · p. 2, Figure 1 · p. 3
- Exploration is constrained to mAP ≥65% on VOC 2007, computational cost ≤5B operations, and 8-bit weight precision.Source: §2.2 · p. 3
3 YOLO Nano Architectural Design
YOLO Nano combines residual PEP and EP modules with fully-connected attention and heterogeneous module designs. The paper attributes reduced complexity with preserved expressiveness to these architectural choices.
3.1 Residual Projection-Expansion-Projection Macroarchitecture
- Residual PEP modules project channels down, expand them, apply depth-wise convolution, then project down again, reducing complexity while preserving expressiveness.Source: §3.1 · pp. 3–4
3.2 Fully-connected Attention Macroarchitecture
- Machine-driven exploration strategically introduces lightweight FCA, whose two fully connected layers learn channel interdependencies and produce modulation weights for channel-wise recalibration.The paper says FCA helps balance reduced complexity with model expressiveness.Source: §3.2 · p. 4
3.3 Macroarchitecture and Microarchitecture Heterogeneity
- The architecture mixes PEP, EP, FCA, and individual 3×3 and 1×1 convolution layers, with unique microarchitectures tailored to balance complexity and expressiveness.Source: §3.3 · p. 4
4 Experimental Results and Discussion
On VOC 2007, YOLO Nano is smaller than Tiny YOLOv2 and Tiny YOLOv3 while achieving higher mAP. Tests on a Jetson AGX Xavier report inference speed and power efficiency at two power budgets.
4 Experimental Results and Discussion
- 4.0MB, >15.1× and >8.3× smaller than Tiny YOLOv2 and Tiny YOLOv3; 69.1% VOC 2007 mAP, ~12% and ~10.7% higher, respectively.Table 1 reports object detection accuracy on the VOC 2007 test set.Source: Table 1 · p. 4, §4 · p. 4
- At 15W and 30W, YOLO Nano achieved ~26.9 FPS and ~48.2 FPS, with power efficiencies of ~1.97 and ~1.61 images/sec/watt, respectively, on Jetson AGX Xavier.Source: §4 · p. 4