# CSL-YOLO: A New Lightweight Object Detection System for Edge Computing

**Authors:** Yu-Ming Zhang, Chun-Chieh Lee, Jun-Wei Hsieh, Kuo-Chin Fan  
**arXiv:** [2107.04829](https://arxiv.org/abs/2107.04829)  
**Date:** 2021-07-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.04829v1/

## Key points

- Development of lightweight object detectors is crucial for edge computing due to limited computation resources
- Proposed Cross-Stage Lightweight (CSL) Module as a new lightweight Convolution method
- CSL-Module generates redundant features from cheap operations, reducing computation cost
- Experiments on MS-COCO dataset show CSL-Module approximates Convolution-3x3 fitting ability while reducing computational requirements
- CSL-YOLO, constructed using CSL-Module, achieves better detection performance with only 43% FLOPs and 52% parameters compared to Tiny-YOLOv4
- CSL Module divides input feature-maps into two branches: one generates half redundant feature-maps using a cheap operation, while the other generates necessary feature-maps using a lightweight main operation
- Architecture of CSL-Module illustrated in Figure 2, with t representing the ratio of feature expanding (t=3 in CSL-Bone)
- Comparisons made between CSL Module and other lightweight convolution methods such as VGG16 and GhostVGG16, showing superior performance in terms of MFLOPs and accuracy.

## Summary

The development of lightweight object detectors is crucial for edge computing due to limited computation resources. To address the challenge of reducing computation cost, this paper proposes a new lightweight Convolution method called Cross-Stage Lightweight (CSL) Module. The CSL-Module generates redundant features from cheap operations, specifically by replacing Pointwise Convolution with Depthwise Convolution in the intermediate expansion stage. This replacement significantly reduces the computation cost. To evaluate the effectiveness of the proposed CSL-Module, experiments were conducted on MS-COCO dataset. The results show that the CSL-Module can approximate the fitting ability of Convolution-3x3 while reducing computational requirements. Furthermore, the module is used to construct a lightweight detector called CSL-YOLO. Compared to Tiny-YOLOv4, CSL-YOLO achieves better detection performance with only 43% FLOPs and 52% parameters. In terms of methodology, this paper introduces the Cross-Stage Lightweight (CSL) Module as a solution to generate feature-maps with fewer FLOPs. The module divides input feature-maps into two branches: one branch generates half redundant feature-maps using a cheap operation similar to GhostNet, while the other branch generates necessary feature-maps using a lightweight main operation. These outputs are then concatenated together. The architecture of CSL-Module is illustrated in Figure 2, where t represents the ratio of feature expanding. In CSL-Bone, t is set as 3, while in other cases it is set as 2. To enhance down sampling or attention after expansion block, an SE module or adaptive average pooling can be inserted. Comparisons were made between CSL Module and other lightweight convolution methods such as VGG16 and GhostVGG16. Table 1 shows that CSLVGG16 outperforms these methods in terms of MFLOPs and accuracy. Overall, this paper presents a novel approach to lightweight object detection by introducing the CSL Module and using it to construct CSL YOLO.