Source-linked AI summary
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
Edge detectors need lower computational cost without sacrificing sufficient accuracy, while depth-wise separable convolution can weaken feature representation through channel separation. The paper introduces GSConv, which mixes standard and depth-wise convolution features, and proposes Slim-Neck for efficient detector design. Across comparative experiments, Slim-Neck detectors achieve strong accuracy with smaller size, while GSConv is especially suited to edge devices because of its low computational and memory consumption.
Problem
Edge devices require real-time detection, but large models are computationally costly and depth-wise-separable-convolution designs may not provide sufficient accuracy because channel information is separated.
Method
The paper introduces GSConv, a mixing strategy combining standard and depth-wise convolution features, and builds the Slim-Neck detector design around it.
Results
Slim-Neck detectors achieve the best accuracy with smaller size in comparisons, with improvements especially pronounced for lightweight detectors such as Yolov3/v4-tiny.
Takeaways & Limitations
GSConv is particularly suitable for edge devices because its computational cost is about 50% of standard convolution while its contribution to model learning ability is comparable.
Takeaways & Limitations
GSConv’s advantages become less obvious as device computing power grows, making it more suitable for edge devices.
Abstract
from arXiv · showhide
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
1. Introduction
Lightweight detectors reduce computation with depth-wise separable convolution, but separating channel information weakens feature representation. The paper introduces GSConv and Slim-Neck designs to retain accuracy while improving efficiency for real-time detection.
- Motivation: Depth-wise separable convolution reduces parameters and FLOPs but separates input-channel information during computation.This separation gives it lower computational cost but weaker feature-representation capability than standard convolution.
- Related work: Prior lightweight models restore channel interaction using dense 1×1 convolutions, channel shuffle, or cheap feature-generation operations.These approaches address the channel-separation flaw through different feature-fusion strategies.
- Contribution: GSConv mixes standard- and depth-wise-convolution features so lightweight representations approach standard-convolution effects with lower computational cost.The technique is designed to cooperate with both convolution types rather than relying exclusively on depth-wise separable convolution.
- Contribution: Slim-Neck uses standard convolution in the backbone and GSConv in the neck to preserve accuracy while accelerating inference.The design targets lightweight real-time detector architectures and reduces the negative impact of depth-wise convolution.
- Evaluation: The paper evaluates the proposed designs through comparative experiments and examines widely used improvements for real-time detectors.The stated contributions include verifying different tricks on popular real-time detectors as a reference for object-detection research.
3. Slim-neck architecture design
The Slim-Neck design combines GSConv, VoV-GSCSP, feature-enhancement choices, activation functions, and IoU losses to build an efficient detector neck. Its components target lower computational complexity and inference time while maintaining detection accuracy.
- Slim-Neck solution: The Slim-Neck solution combines GSConv, VoV-GSCSP, activation choices, bounding-box losses, and other detector-improvement tricks.The stated aim is a simple and efficient neck for detectors deployed on edge devices.
- GSConv: The method evaluates shuffle implementations for GSConv, including transposition without additional FLOPs and lower-cost linear operations supported by convolution-capable devices.The transposition approach may be unsupported on some edge devices, whereas the linear-operation alternative is canonical and broadly supported.
- VoV-GSCSP: The GS bottleneck and VoV-GSCSP modules extend GSConv into neck blocks designed to reduce computational complexity and inference time.VoV-GSCSP uses a one-shot aggregation strategy and replaces ordinary bottlenecks in the proposed design.
- Activation and loss choices: EIoU independently penalizes predicted width and height, addressing a CIoU limitation when object width and height differ substantially.The passage states that EIoU bounding-box regression is more exact especially for objects whose width greatly differs from height.
- Activation and loss choices: Mish can provide slightly better model accuracy than Swish, but its exponential and logarithmic operations increase training cost.Ablation experiments compare IoU-loss and activation-function combinations.
4. Experiments and Analysis
Experiments evaluate GSConv, Slim-Neck designs, attention modules, activation and loss choices across pedestrian, general object-detection, traffic, and edge-device settings. The results report accuracy, computational cost, latency, and deployment trade-offs against baseline and lightweight detectors.
- Datasets: The experiments use WiderPerson, VOC07+12, and SODA10M to assess practical, lightweight, and traffic-environment detection performance.WiderPerson contains dense pedestrian scenes with overlap and occlusion; VOC07+12 supports comparisons with lightweight models, while SODA10M tests traffic performance.
- GSConv ablation: GSConv achieves the best outcomes among five convolution methods on VOC07+12 with lower time complexity in the Yolov5n backbone.The comparison includes SC, DSC, ShuffleNet, GhostNet, and GSConv at 1/8, 1/16, and 1/32 down-sample scales.
- Activation and loss functions: Mish with EIoU achieves higher average accuracy, whereas Swish achieves faster speed, with Mish increasing training time by about 29.26%.Table 7 compares Mish and Swish activation functions together with CIoU and EIoU regression losses.
- SNs comparisons: SNs detectors achieve the best accuracy with smaller size, and combining SNs with additional tricks substantially improves lightweight detectors such as Yolov3/v4-tiny.SNs-Yolo-tiny models are also evaluated on the Jetson Nano, while SODA10M compares speed and accuracy among five state-of-the-art detectors.
- Ablation studies: Table 8 ablates SNs, attention modules, and activation and IoU-loss functions, while noting that equal latency can yield different FPS because of postprocessing time.The table compares variants using different VoV-GSCSP modules and reports the effects of these design choices on VOC07+12.
5. Discussion
The paper positions GSConv as a flexible lightweight design option for CNN-based real-time detectors, especially on edge devices. Its Slim-Neck designs target computational efficiency while preserving detection accuracy, though their advantages diminish with greater device computing power.
- GSConv offers a lightweight design option for CNN-based vision models, which remain preferred for industrial applications because Transformer-based detectors face latency challenges.
- GSConv is more suitable for edge computing because its advantages become less obvious as device computing power grows.The paper also describes promising tests in traffic assistance systems and proposes agricultural edge-vision applications.
- The GSConv can support additional branches for coordinate encoding, large receptive fields, or spatial attention-like behavior.The auxiliary depth-wise separable convolution branch can reach a 17 × 17 kernel without overfitting in reported tests.
- The Slim-Neck designs reduce computational cost by 7.2%~32.7% while improving accuracy by 1.2%~35.3% in the tested detectors.The reported results outperform state-of-the-art detectors.
A. The comparisons of the FLOPs and parameters amount between the DSC, SC and GSConv
The section compares conventional convolution (SC), depth-wise separable convolution (DSC), and GSConv through parameter counts and computational-cost ratios. DSC uses fewer parameters and lower computational cost than SC, while GSConv is evaluated relative to SC.
- Parameter and FLOPs comparison: DSC requires fewer parameters than a conventional convolutional layer by separating spatial convolution from 1×1 channel mixing.The conventional layer uses C1 × K1 × K2 × C2 parameters, whereas DSC uses a depth-wise term plus 1×1 convolution.
- Parameter and FLOPs comparison: DSC has lower calculation cost than conventional convolution, with cost expressed using feature-map width W, height H, channels C1 and C2, and kernel size K1×K2.The section defines separate computational-cost expressions for SC and DSC and states that DSC’s cost is much lower.
- Parameter and FLOPs comparison: The ratios ratiop and ratioc denote DSC-to-SC parameter and computational-cost ratios, respectively.The text uses these ratios to quantify the reduction achieved by DSC relative to conventional convolution.
- Parameter and FLOPs comparison: GSConv is likewise assessed through its computational-cost ratio relative to conventional convolution.The displayed GSConv ratio is introduced after the DSC comparison, but the supplied formula text is fragmented.
B Comparison of remote sensing image detection
The experiments compare SNs-enhanced and original YOLOv4 detectors for remote-sensing images and evaluate lightweight detectors in low-light traffic scenes. Results are presented through detection images and field-video frames, including inference on a Jetson Nano.
- Remote sensing image detection: SNs-Yolov4 and Yolov4 are trained with the same hyperparameters on DOTA1.0 to compare their ability to detect small objects.Training uses an A40, and Figures a and b show the respective test results.
- Low-light field evaluation: A 20-second dash-cam video captured at night in low light is used to test the approach, with four frames shown for intuitive comparison.The evaluation focuses on field traffic scenes under low-light conditions.
- Remote sensing image detection: Figure a presents the SNs-Yolov4 detection result, while Figure b presents the original Yolov4 detection result.The figures provide a direct visual comparison of the two detectors.
- Low-light field evaluation: Figure c shows visual results from SNs-Yolo in field traffic running on a Jetson Nano.The figure documents deployment-oriented inference in a traffic setting.