This website requires JavaScript.

DATE: Dual Assignment for End-to-End Fully Convolutional Object Detection

Yiqun ChenQiang ChenQinghao HuJian Cheng
Nov 2022
摘要
Fully convolutional detectors discard the one-to-many assignment and adopt aone-to-one assigning strategy to achieve end-to-end detection but suffer fromthe slow convergence issue. In this paper, we revisit these two assignmentmethods and find that bringing one-to-many assignment back to end-to-end fullyconvolutional detectors helps with model convergence. Based on thisobservation, we propose {\em \textbf{D}ual \textbf{A}ssignment} for end-to-endfully convolutional de\textbf{TE}ction (DATE). Our method constructs twobranches with one-to-many and one-to-one assignment during training and speedsup the convergence of the one-to-one assignment branch by providing moresupervision signals. DATE only uses the branch with the one-to-one matchingstrategy for model inference, which doesn't bring inference overhead.Experimental results show that Dual Assignment gives nontrivial improvementsand speeds up model convergence upon OneNet and DeFCN. Code:https://github.com/YiqunChen1999/date.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答