This website requires JavaScript.

FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling

Hao LuWenze LiuHongtao FuZhiguo Cao
Jul 2022
摘要
We consider the problem of task-agnostic feature upsampling in denseprediction where an upsampling operator is required to facilitate bothregion-sensitive tasks like semantic segmentation and detail-sensitive taskssuch as image matting. Existing upsampling operators often can work well ineither type of the tasks, but not both. In this work, we present FADE, a novel,plug-and-play, and task-agnostic upsampling operator. FADE benefits from threedesign choices: i) considering encoder and decoder features jointly inupsampling kernel generation; ii) an efficient semi-shift convolutionaloperator that enables granular control over how each feature point contributesto upsampling kernels; iii) a decoder-dependent gating mechanism for enhanceddetail delineation. We first study the upsampling properties of FADE on toydata and then evaluate it on large-scale semantic segmentation and imagematting. In particular, FADE reveals its effectiveness and task-agnosticcharacteristic by consistently outperforming recent dynamic upsamplingoperators in different tasks. It also generalizes well across convolutional andtransformer architectures with little computational overhead. Our workadditionally provides thoughtful insights on what makes for task-agnosticupsampling. Code is available at: http://lnkiy.in/fade_in
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