Improve yolov5 with gsconv+slim neck to maximize performance!

Paper title ：Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles

Object detection is an arduous downstream task in computer vision . For vehicle edge computing platform , Large models are difficult to meet the requirements of real-time detection . and , A lightweight model constructed from a large number of deep separable convolutions cannot achieve sufficient accuracy . So this paper introduces a new method GSConv To reduce the complexity of the model and maintain accuracy .GSConv It can better balance the accuracy and speed of the model . also , Provides a design paradigm ,Slim-Neck, To achieve higher computational cost-effectiveness of the detector . In the experiment , Compared with the original network , The most advanced results are obtained by this method （ for example ,SODA10M stay Tesla T4 In order to ~100FPS The speed of 70.9% mAP0.5）.

The main contributions can be summarized as follows ：

1. Introduced a new method  GSConv  Instead of  SC  operation . This method makes the output of convolution calculation as close as possible to  SC, At the same time, reduce the cost of Computing ;
2. It provides a new design paradigm for the detector architecture of autonomous vehicle , I.e. with standard  Backbone  Of  Slim-Neck  Design ;
3. Verified the difference Trick The effectiveness of the , It can be used as a reference for research in this field .