ssd

SSD是一种非常优秀的one-stage目标检测方法，one-stage算法就是目标检测和分类是同时完成的，其主要思路是利用CNN提取特征后，均匀地在图片的不同位置进行密集抽样，抽样时可以采用不同尺度和长宽比，物体分类与预测框的回归同时进行，整个过程只需要一步，所以其优势是速度快。
但是均匀的密集采样的一个重要缺点是训练比较困难，这主要是因为正样本与负样本（背景）极其不均衡（参见Focal Loss），导致模型准确度稍低。
SSD的英文全名是Single Shot MultiBox Detector，Single shot说明SSD算法属于one-stage方法，MultiBox说明SSD算法基于多框预测。
————————————————

版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。

原文链接：https://blog.csdn.net/weixin_44791964/article/details/104981486

SSD is an outstanding one-stage object detection algorithm. One-stage detection algorithms complete object detection and classification simultaneously. The core idea of SSD is to perform uniform dense sampling at different positions on the input image after extracting features via CNN, where multiple scales and aspect ratios can be adopted. It conducts object classification and bounding box regression at the same time, with the entire process completed in one single step, hence its advantage of fast inference speed.

However, a major drawback of uniform dense sampling is the challenge in training, which is primarily attributed to the severe imbalance between positive and negative samples (background), as referenced in Focal Loss, resulting in slightly reduced model accuracy.

The full English name of SSD is Single Shot MultiBox Detector. The term "Single Shot" indicates that SSD falls into the category of one-stage object detection algorithms, while "MultiBox" means that SSD is based on the multi-box prediction framework.

————————————————
Copyright Notice: This article is an original work by the blogger, licensed under the CC 4.0 BY-SA copyright agreement. Please attach the original source link and this notice when reposting.
Original link: https://blog.csdn.net/weixin_44791964/article/details/104981486

搜集汇总

背景与挑战

背景概述

该数据集涉及SSD（Single Shot MultiBox Detector）目标检测方法，这是一种单阶段算法，通过卷积神经网络提取特征并密集抽样，同步实现物体分类与边界框回归，具有速度快的特点，但训练中可能因样本不均衡而面临挑战。