ST3D(未完待续)

Posted 2023-03-09 Keep--Silent

ST3D: Self-training for Unsupervised Domain Adaptation on 3D Object Detection
三维物体检测中弱监督领域自适应的自训练
3D detectors developed on one specific domain (i.e. source domain) might not generalize well to novel testing domains (i.e. target domains) due to unavoidable domain-shifts arising from different types of 3D sensors, weather conditions and geographical locations, etc
由于不同类型的3D传感器、天气条件和地理位置等导致的不可避免的领域变化，在一个特定领域（即源领域）开发的3D检测器可能无法很好地推广到新的测试领域（即目标领域）。

Therefore, approaches to effectively adapting 3D detector trained on labeled source domain to a new unlabeled target domain is highly demanded in practical applications.

因此，在实际应用中，有效地适应在标记源域上训练的3D检测器到一个新的未标记目标域是非常需要的方法

ST3D

• 总体
• • Random Object Scaling (ROS)
  • • 目的
    • 步骤
  • Pseudo label Generation with Quality-aware Triplet Memory Bank (QTMB)
  • • 目的
    • 组成
    • 步骤
    • • Proxy-pseudo Label $[{\hat{L}}_i^t{]}_k{}_{i=1}^{n_t}$
  • Curriculum Data Augmentation (CDA)
• Algorithm
• • 符号说明
  • 输入：源域的点云和标签$(P_i^s,L_i^s){}_{i=1}^{n_s}$，目标域的点云$P_i^t{}_{i=1}^{n_t}$
  • 输出：目标域的三维目标检测模型。

总体

Random Object Scaling (ROS)

随机目标缩放策略

目的

overcome the bias in object size on the labeled source domain
克服了标记源域上对象大小的偏差。

ROS可以有效地模拟具有不同对象大小的对象，以解决大小偏差，从而有助于训练大小鲁棒的检测器，为后续的自我训练产生更精确的初始伪盒

步骤

给定3D bounding box with size大小 $(l,w,h)$, center中心 $(c_x,c_y,c_z)$ and heading angle $\theta$ , random scale factors随机规模因子 $(r_l,r_w,r_h)$
有$n_p$个点, 分别是$p_i{}_{i=1}^{n_p}$, 其坐标是$(p_i^x,p_i^y,p_i^z)$

First, we transform the points to the local coordinate system of the box along its length, width and height dimensions via

$(p_i^l,p_i^w,p_i^h)=(p_i^x-c_x,p_i^y-c_y,p_i^z-c_z)\cdot R$
$R=\left[\begin{array}{ccc}\cos \theta & -\sin \theta & 0\\ \sin \theta & \cos \theta & 0\\ 0& 0& 1\end{array}\right],$
where · is matrix multiplication.
通过该公式将这些点转化为局部坐标系

Second, to derive the scaled object, the point coordinates inside the box are scaled to be $(r_l{p}_i^l,r_w{p}^{w}i,r_h{p}_i^h)$ with object size $(r_{l}l,r_{w}w,r_{h}h)$.
框内点坐标被缩放为$(r_l{p}_i^l,r_w{p}^{w}i,r_h{p}_i^h)$

Third, to derive the augmented data $p_i^{aug}{}_{i=1}^{n_p}$, the points inside the scaled box are transformed back to the ego-car coordinate system and shifted to the center $(c_x,c_y,c$

前端调试（未完待续）debugger

未完待续

ant使用(未完待续)

漫谈回溯(未完待续)

未完待续