Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Slim-neck by GSConv: A lightweight-design for real-time detector architectures

Hulin Li, Jun Li, Hanbing Wei, Zheng Liu, Zhenfei Zhan, Qiliang Ren

TL;DR

A new lightweight convolutional technique, GSConv, is introduced to lighten the model but maintain the accuracy, and the real-time detectors of ameliorated by the SNs obtain the state-of-the-art results.

Abstract

Real-time object detection is significant for industrial and research fields. On edge devices, a giant model is difficult to achieve the real-time detecting requirement and a lightweight model built from a large number of the depth-wise separable convolutional could not achieve the sufficient accuracy. We introduce a new lightweight convolutional technique, GSConv, to lighten the model but maintain the accuracy. The GSConv accomplishes an excellent trade-off between the accuracy and speed. Furthermore, we provide a design suggestion based on the GSConv, Slim-Neck (SNs), to achieve a higher computational cost-effectiveness of the real-time detectors. The effectiveness of the SNs was robustly demonstrated in over twenty sets comparative experiments. In particular, the real-time detectors of ameliorated by the SNs obtain the state-of-the-art (70.9% AP50 for the SODA10M at a speed of ~ 100FPS on a Tesla T4) compared with the baselines. Code is available at https://github.com/alanli1997/slim-neck-by-gsconv

Slim-neck by GSConv: A lightweight-design for real-time detector architectures

TL;DR

A new lightweight convolutional technique, GSConv, is introduced to lighten the model but maintain the accuracy, and the real-time detectors of ameliorated by the SNs obtain the state-of-the-art results.

Abstract

Real-time object detection is significant for industrial and research fields. On edge devices, a giant model is difficult to achieve the real-time detecting requirement and a lightweight model built from a large number of the depth-wise separable convolutional could not achieve the sufficient accuracy. We introduce a new lightweight convolutional technique, GSConv, to lighten the model but maintain the accuracy. The GSConv accomplishes an excellent trade-off between the accuracy and speed. Furthermore, we provide a design suggestion based on the GSConv, Slim-Neck (SNs), to achieve a higher computational cost-effectiveness of the real-time detectors. The effectiveness of the SNs was robustly demonstrated in over twenty sets comparative experiments. In particular, the real-time detectors of ameliorated by the SNs obtain the state-of-the-art (70.9% AP50 for the SODA10M at a speed of ~ 100FPS on a Tesla T4) compared with the baselines. Code is available at https://github.com/alanli1997/slim-neck-by-gsconv
Paper Structure (19 sections, 6 equations, 7 figures, 10 tables)

This paper contains 19 sections, 6 equations, 7 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Slim-neck architecture design
  4. GSConv
  5. VoV-GSCSP block
  6. Free improvement tricks for lightweight models
  7. Loss and activation
  8. IoU loss function choosing
  9. Activation function choosing
  10. The SNs for real-time detectors
  11. Experiments and Analysis
  12. Data Sets
  13. Ablation Studies
  14. Comparisons between the SNs detectors and baselines
  15. Discussion
  16. ...and 4 more sections

Figures (7)

  • Figure 1: The calculation process of the (a) SC and the (b) DSC. The SC is the channel-dense convolutional computation and the DSC is the channel-sparse.
  • Figure 2: The structure of the GSConv. The "SC" marked in orange means the standard convolutional-2D layer with a batch normalization-2D layer and an activation layer. The "DSC" marked in blue means the depth-wise convolutional-2D layer with a batch normalization-2D layer and an activation layer.
  • Figure 3: Three type features of the different basic convolutional of the scale 1/4 (P2) of the Yolov5n.
  • Figure 4: The structures of the (a) GS bottleneck module and the (b), (c), (d) VoV-GSCSP${ }_{1,2,3}$ modules.
  • Figure 5: Comparisons of the SNs-Yolo and baselines 6. Conclusion
  • ...and 2 more figures