論文標題:MEnsA: Mix-up Ensemble Average for Unsupervised Multi Target Domain Adaptation on 3D Point Clouds
論文作者:Ashish Sinha, Jonghyun Choi
論文來源:2023 CVPR
論文地址:download
論文程式碼:download
視屏講解:click
單目標域和多目標域
單目標域和多目標域的差異:
Mixup 模組:
$F_{i}^{m}=\lambda F_{s}+(1-\lambda) F_{T_{i}} \quad\quad(1)$
$L_{i}^{m}=\lambda L_{s}+(1-\lambda) L_{T_{i}} \quad\quad(2)$
線性差值的好處:
基線:【多目標域場景下】
本文:單源域 和 多目標域整合線性差值;
總損失:
$\mathcal{L}=\log \left(\sum\left(e^{\gamma\left(\mathcal{L}_{c l s}+\eta \mathcal{L}_{d c}+\zeta \mathcal{L}_{a d v}\right)}\right)\right) / \gamma \quad\quad(4)$
其中:
源域分類損失: $\mathcal{L}_{c l s} =\mathcal{L}_{C E}\left(C\left(F_{s}\right), y_{s}\right) \quad\quad(4)$
單源域單目標域鑑別損失:$\mathcal{L}_{d c} =\mathcal{L}_{C E}\left(D\left(F_{s}\right), L_{s}\right)+\mathcal{L}_{C E}\left(D\left(F_{T_{i}}, L_{T_{i}}\right)\right) \quad\quad(5)$
對抗損失:$\mathcal{L}_{a d v} =\lambda_{1} \mathcal{L}_{m m d}+\lambda_{2} \mathcal{L}_{d c}+\lambda_{3} \mathcal{L}_{\text {mixup }} \quad\quad(6)$
關於對抗損失:
MMD 損失:$\mathcal{L}_{m m d}=\mathcal{L}_{r b f}\left(C\left(F_{s}\right), F_{T_{i}}, \sigma\right) \quad\quad(7)$
線性差值域鑑別損失:$\mathcal{L}_{\text {mixup }}=\mathcal{L}_{C E}\left(D\left(F_{m}^{M}\right), L_{i}^{m}\right) \quad\quad(8)$
Note:
線性差值:
$F_{m}^{\text {factor }}=\lambda F_{s}+\sum_{i=1}^{n} \frac{1-\lambda}{n} F_{T_{i}}$
$F_{m}^{\text {concat }}=\left[\lambda F_{s}, \frac{1-\lambda}{n} F_{T_{1}}, \ldots, \frac{1-\lambda}{n} F_{T_{n}}\right]$
$.L_{m}^{\text {concat }}=[\lambda, 2 \frac{1-\lambda}{n}, \ldots, N \frac{1-\lambda}{n}]$
$F_{m}^{T}=\lambda F_{T_{1}}+(1-\lambda) F_{T_{2}}$
$L_{m}^{T}=\lambda L_{T_{1}}+(1-\lambda) L_{T_{2}} $
因上求緣,果上努力~~~~ 作者:VX賬號X466550,轉載請註明原文連結:https://www.cnblogs.com/BlairGrowing/p/17348882.html