PaperlayerView exact revision on arXiv

Source-linked AI summary

EdgeYOLO: An Edge-Real-Time Object Detector

Shihan Liu, Junlin Zha, Jian Sun, Zhuo Li, Gang Wang

arXiv:2302.07483v1cs.CV

TL;DR

EdgeYOLO addresses the challenge of accurate real-time object detection on edge devices. It proposes an anchor-free detector with training changes for richer data and small-object detection, and reports real-time edge performance with improved small-object results.

  • Problem

    YOLO models with state-of-the-art performance often have unsatisfactory FPS on edge devices, while lighter designs can sacrifice precision.

    Source: §1 · p. 1

  • Method

    EdgeYOLO uses an anchor-free detector, richer data augmentation, and a redesigned loss to improve small-object detection.

    Source: §1 · p. 1, §2.5 · p. 3

  • Results

    EdgeYOLO runs on edge devices in real time with high accuracy, and small-object performance increases by 2.2% AP.

    Source: §4.2 · p. 5, §5 · p. 6

  • Takeaways & Limitations

    The anchor-free structure reduces design and computational complexity and is described as more deployment-friendly for edge devices.

    Source: §5 · p. 6

Abstract

from arXiv · show

This paper proposes an efficient, low-complexity and anchor-free object detector based on the state-of-the-art YOLO framework, which can be implemented in real time on edge computing platforms. We develop an enhanced data augmentation method to effectively suppress overfitting during training, and design a hybrid random loss function to improve the detection accuracy of small objects. Inspired by FCOS, a lighter and more efficient decoupled head is proposed, and its inference speed can be improved with little loss of precision. Our baseline model can reach the accuracy of 50.6% AP50:95 and 69.8% AP50 in MS COCO2017 dataset, 26.4% AP50:95 and 44.8% AP50 in VisDrone2019-DET dataset, and it meets real-time requirements (FPS>=30) on edge-computing device Nvidia Jetson AGX Xavier. We also designed lighter models with less parameters for edge computing devices with lower computing power, which also show better performances. Our source code, hyper-parameters and model weights are all available at https://github.com/LSH9832/edgeyolo.

1 Introduction

Object detection balances accuracy with real-time inference, but high-performing detectors often run too slowly on edge devices. EdgeYOLO targets this gap with an anchor-free detector and training and model-design changes intended to preserve accuracy while enabling edge deployment.

1 Introduction

  • Two-stage detectors can perform somewhat better on common datasets, but their framework limits make real-time operation difficult, while one-stage detectors balance speed and performance.The paper identifies YOLO as a popular one-stage framework.Source: §1 · p. 1
  • State-of-the-art YOLO models are often tested on powerful workstations yet can have unsatisfactory FPS on edge devices, motivating lighter designs that trade some precision for speed.The paper aims to retain decent precision while running in real time on edge devices.Source: §1 · p. 1
  • EdgeYOLO is an anchor-free detector designed for real-time edge-device use, with a reported 50.6% AP on MS COCO2017.The introduction also proposes richer data augmentation, reparameterized structures to reduce inference time, and a loss function aimed at small-object precision.Source: §1 · p. 1

2 Related Work

The related work motivates anchor-free detection for edge latency, flexible augmentation for datasets with sparse labels, and model and head designs that balance speed with precision. For small objects, the paper combines scaling and stitching augmentation with a redesigned loss.

2.1 Anchor-free Object Detector

  • Anchor-free detection can reduce edge-device post-processing time by more than half relative to anchor-based frameworks, motivating the paper’s choice of strategy.Anchor-based YOLO typically assigns three anchors to each grid cell; the text says post-processing latency is nearly proportional to anchors per cell.Source: §2.1 · p. 2

2.2 Data Augmentation

  • The paper’s combined augmentation is intended to increase input-data richness and validity, addressing overfitting and the risk of augmented images with too few valid objects.Common augmentation methods mix pixel information from multiple pictures; the proposed method is described as more flexible and powerful.Source: §2.2 · p. 2

2.3 Model Reduction

  • Combining lossy and lossless model reduction yields models of different sizes for edge devices with different computing power while speeding inference.Lossy reduction shrinks networks by reducing layers or channels, while lossless reduction re-parameterizes multiple branches into a streamlined equivalent module.Source: §2.3 · pp. 2–3

2.4 Decoupled Regression

  • A lighter decoupled head addresses the extra inference cost of decoupling location, confidence, and category regression, whose relationships are described as insufficiently close.The paper notes that decoupled heads can improve results and accelerate loss convergence, but add inference costs.Source: §2.4 · p. 3

2.5 Small Object Detecting Optimization

  • Small objects provide less pixel information and often have lower detection precision; the paper uses scaling and stitching augmentation plus a redesigned loss rather than replication augmentation.The text identifies scale and background mismatch as problems with replication augmentation and says the selected methods improve detection of small and medium objects and overall precision.Source: §2.5 · p. 3

3 Approach

The approach addresses augmentation cases that can produce images without valid objects, then combines enhanced Mosaic+Mixup, a lighter decoupled head, and staged losses. These designs target richer training images, faster inference, and improved small-object precision.

3.1 Enhanced-Mosaic & Mixup

  • The enhanced Mosaic+Mixup method combines groups of Mosaic-processed images with a final image, ensuring sufficient effective information while alleviating overfitting.The number of Mosaic groups can be set according to the average label richness of images in the dataset.Source: §3.1 · p. 3

3.2 Lite-Decoupled Head

  • The lighter decoupled head reduces channels and convolutional layers, adds implicit representation layers, and re-parameterizes them to lower inference costs.The box and confidence regression layers are also merged for high-parallelism inference; the figure caption reports faster inference with little precision loss.Source: §3.2 · p. 4, Figure 4 · p. 4

3.3 Staged Loss Function

  • Training has three stages: the second replaces classification and object losses with Hybrid-Random Loss, and the third closes augmentation and switches to L1 regulation and cIOU.In the second stage, the random loss uses predictions, ground truth, and a random value between 0 and 1; the authors say it balances small-object and total precision.Source: §3.3 · p. 4

4 Experiments

Experiments evaluate EdgeYOLO on COCO and VisDrone, compare it with other detectors, and test design choices and deployment adjustments. Results report real-time edge-device detection, small-object gains, and improvements from selected head and loss configurations.

4.1 Implementation Details

  • Experiments use COCO as the main training benchmark and VisDrone specifically to test small-object detection; inference FPS is measured on Jetson AGX Xavier in FP16 with TensorRT.Training uses ELAN-Darknet, SGD, and a 300-epoch schedule.Source: §4.1 · pp. 4–5

4.2 Results & Comparison

  • EdgeYOLO is reported as a high-precision real-time Jetson AGX Xavier detector with fewer parameters, a 2.2% AP increase on small objects, and a VisDrone lead over larger models.The COCO comparison also reports precision exceeding some non-real-time models.Source: Table 1 · p. 5, §4.2 · p. 5, Table 2 · p. 5

4.3 Ablation Study

  • A lightweight decoupled head improves precision without much real-time performance loss, segmentation labels add 2%–3% AP during stable augmentation but only about 0.04% final AP when omitted, and HR loss with cIOU performs better in later training.With HR loss, gIOU raises small-object precision by 0.3% AP but lowers large-object precision by 0.7% AP relative to cIOU.Source: §4.3 · pp. 5–6, Table 3 · p. 5, Table 4 · p. 6

4.4 Tricks for Edge Computing Devices

  • For edge deployment, 640×480 or 640×384 inputs can improve inference speed without accuracy loss, while splitting preprocessing, model input, and postprocessing across processes and threads increases FPS by about 8%–14%.The input sizes correspond to common 4:3 and 16:9 video-stream aspect ratios.Source: §4.4 · p. 6, Table 5 · p. 6

5 Conclusion

The conclusion presents EdgeYOLO as an accurate, real-time anchor-free detector for edge devices, with improved small-object detection. It identifies further small-object accuracy improvements and efficient optimization as future work.

5 Conclusion

  • EdgeYOLO is reported to run in real time on edge devices with high accuracy and improved small-object detection.The conclusion says its anchor-free structure reduces design and computational complexity and is friendlier to edge deployment.Source: §5 · p. 6
  • The authors identify extending the framework to pixel-level recognition tasks such as instance segmentation as a possible direction.Source: §5 · p. 6
  • Future work will further improve small-object detection accuracy and explore efficient optimizations.Source: §5 · p. 6

Related papers

Loading 2302.07483v1…