人工智能培训

搜索

深度学习论文:深度完成的深度系数(Depth Coefficients for Depth Completion)

[复制链接]
fanjz 发表于 2019-3-15 13:20:04 | 显示全部楼层 |阅读模式
fanjz 2019-3-15 13:20:04 159 0 显示全部楼层
深度学习论文:深度完成的深度系数(Depth Coefficients for Depth Completion)深度完成涉及从稀疏深度测量估计密集深度图像,通常由彩色图像引导。虽然线性上采样是直接向前的,但是它导致包括深度像素的伪像在对象之间的不连续处的空白空间中被插入。当前的方法使用深度网络来上采样并“完成”缺失的深度像素。然而,对象之间的深度拖尾仍然是一个挑战。我们提出了一种称为深度系数(DC)的深度的新表示来解决这个问题。它使卷积更容易避免对象间深度混合。我们还表明,标准均方误差(MSE)损失函数可以促进深度混合,因此建议使用DC的交叉熵损失。通过对基准的定量和定性评估,我们表明,通过我们的DC表示和交叉熵损失来切换稀疏深度输入和MSE损失是提高深度完成性能和减少像素深度混合的简单方法,这导致改进的基于深度的物体检测。
Depth completion involves estimating a dense depth image from sparse depthmeasurements, often guided by a color image.While linear upsampling isstraight forward, it results in artifacts including depth pixels beinginterpolated in empty space across discontinuities between objects.Currentmethods use deep networks to upsample and "complete" the missing depth pixels.Nevertheless, depth smearing between objects remains a challenge.We propose anew representation for depth called Depth Coefficients (DC) to address thisproblem.It enables convolutions to more easily avoid inter-object depthmixing.We also show that the standard Mean Squared Error (MSE) loss functioncan promote depth mixing, and thus propose instead to use cross-entropy lossfor DC.With quantitative and qualitative evaluation on benchmarks, we showthat switching out sparse depth input and MSE loss with our DC representationand cross-entropy loss is a simple way to improve depth completion performance,and reduce pixel depth mixing, which leads to improved depth-based objectdetection.深度学习论文:深度完成的深度系数(Depth Coefficients for Depth Completion) GGcSZOGOcPZcv5k5.jpg
URL地址:https://arxiv.org/abs/1903.05421     ----pdf下载地址:https://arxiv.org/pdf/1903.05421    ----深度学习论文:深度完成的深度系数(Depth Coefficients for Depth Completion)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

fanjz当前离线
新手上路

查看:159 | 回复:0

快速回复 返回顶部 返回列表