Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
three_interpolate 是 PointNet++ 点云处理网络中的核心算子,用于基于3个最近邻的加权线性特征插值。给定已知特征点的特征描述符、待插值点的3个最近邻索引和对应权重,对每个待插值点的每个特征通道,计算3个最近邻特征的加权和。
数学公式:
对于每个 batch b、通道 c、待插值点 n:
y[b, c, n] = Σ(k=0..2) weight[b, n, k] * features[b, c, idx[b, n, k]]
即: y[b, c, n] = weight[b, n, 0] * features[b, c, idx[b, n, 0]] + weight[b, n, 1] * features[b, c, idx[b, n, 1]] + weight[b, n, 2] * features[b, c, idx[b, n, 2]]
长尾算子
/assign
Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
Backgroud(背景信息)
three_interpolate 是 PointNet++ 点云处理网络中的核心算子,用于基于3个最近邻的加权线性特征插值。给定已知特征点的特征描述符、待插值点的3个最近邻索引和对应权重,对每个待插值点的每个特征通道,计算3个最近邻特征的加权和。
数学公式:
对于每个 batch b、通道 c、待插值点 n:
y[b, c, n] = Σ(k=0..2) weight[b, n, k] * features[b, c, idx[b, n, k]]
即:
y[b, c, n] = weight[b, n, 0] * features[b, c, idx[b, n, 0]]
+ weight[b, n, 1] * features[b, c, idx[b, n, 1]]
+ weight[b, n, 2] * features[b, c, idx[b, n, 2]]
Origin(信息来源)
长尾算子
Benefit / Necessity (价值/作用)
Design(设计方案)