Deep Hough Voting for 3D Object Detection in Point Clouds

(https://link.zhihu.com/?target=https://github.com/facebookresearch/votenet)

### 问题与灵感

整个方法based PointNet++

#### 问题

  1. 3D目标检测的中心往往存在3D点云的内部不存在的点上，直接对点云的表面去预测中心所在的点效果可能不佳

#### 灵感

  1. 通过霍夫投票的方式，对3D目标进行检测。
  2. 霍夫投票对于离散的点具有更高的可计算性

霍夫变换相关工作

>
>
>   1. Jan Knopp, Mukta Prasad, and Luc Van Gool. Orientation invariant 3d object classification using hough transform based methods. In Proceedings of the ACM workshop on 3D object retrieval, pages 15–20. ACM, 2010. 2
>   2. Jan Knopp, Mukta Prasad, and Luc Van Gool. Scene cut: Class-specific object detection and segmentation in 3d scenes. In 2011 International Conference on 3D Imaging, Modeling, Processing, Visualization and Transmission, pages 180–187. IEEE, 2011. 2
>   3. Alexander Velizhev, Roman Shapovalov, and Konrad Schindler. Implicit shape models for object detection in 3d point clouds. In International Society of Photogrammetry and Remote Sensing Congress, volume 2, 2012. 2
>   4. Oliver JWoodford, Minh-Tri Pham, Atsuto Maki, Frank Perbet, and Bj¨orn Stenger. Demisting the hough transform for 3d shape recognition and registration. International Journal of Computer Vision, 106(3):332–341, 2014. 2
>

### 论文架构
![在这里插入图片描述](https://ad.itadn.com/c/weblog/blog-img/images/2025-01-25/KzaEZv051p6QPqLJV3HYlhbOCu97.png)

#### Learning to Vote in Point Clouds

##### generate vote

以N个具有3D坐标的点作为输入 -> 产生M种子点。

  * input：N×3
  * backbone：PointNet++ (several set-abstraction layers + feature propagation layers)
  * output: $s_i=[x_i:f_i]$, M×(3+C)，其中 3 每个种子的3D坐标，C高维度特征.

##### Hough voting with deep networks

以vote 作为输入，计算每个vote中到中心点的偏移量

  * input： $s_i=[x_i:f_i]$
  * multi-layer MLP netword: FN+BN+ReLU
  * -output: $\Delta$$x_i$ 和 $\Delta$$f_i$  
根据偏移量，得到候选中心点 $v_i=[y_i:g_i]$  
$y_i=x_i+\Delta x_i$ $g_i = f_i + \Delta f_i$

#### Object Proposal and Classification from Votes

##### Vote clustering through sampling and grouping

依据$v_i$ 得到K个聚类中心

$C_k={v_i^k | ||v_i-v_ik|| \leq r }$

##### Proposal and classification from vote clusters

input $C = {w_i}$ vote cluster $w_i$ 中心

将个点的远点调整到$w_i$  
$z_i'=(z_i-z_j)/r$

![在这里插入图片描述](https://ad.itadn.com/c/weblog/blog-img/images/2025-01-25/27N9mM8TnYIzWeGpHCjSJfdxy4gk.png)

通过两层的MLP计算目标检测信息

output：128：5+2NH+4NS+NC

    * 2 objectness scores
    * 3 center regression values
    * 2NH numbers for heading regression (NH heading bins)
    * 4NS numbers for box size regression (NS box anchors)
    * NC numbers for semantic classification）

### 结果分析
![在这里插入图片描述](https://ad.itadn.com/c/weblog/blog-img/images/2025-01-25/H9wZVFU8z42uJ6Y7OfkRm1Tt5lvX.png)
![在这里插入图片描述](https://ad.itadn.com/c/weblog/blog-img/images/2025-01-25/9y2AJbMpi5SjFuD1EWQNdRZql64U.png)