MV3D -- 3D proposal 网络

3D proposal 网络

实际上3D proposal network就是2D proposal的升维，从实现来看，本质思想完全没有变化。然而实际上第一步得到的3D proposal是通过M+2个投影后的鸟瞰图得到的，这一部分理解起来较难。

• Given a bird’s eye view map. the network generates 3D box proposals from a set of 3D prior boxes. Each 3D boxis parameterized by (x, y, z, l,w,h), which are the center and size (in meters) of the 3D box in LIDAR coordinate system. For each 3D prior box, the corresponding bird’s eye view anchor (xbv, ybv, lbv,wbv) can be obtained by dis- cretizing (x, y, l,w). We design N 3D prior boxes by clus- tering ground truth object sizes in the training set. In the case of car detection, (l,w) of prior boxes takes values in
  {(3.9, 1.6), (1.0, 0.6)}, and the height h is set to 1.56m. By rotating the bird’s eye view anchors 90 degrees, we obtain
  N = 4 prior boxes. (x, y) is the varying positions in the bird’s eye view feature map, and z can be computed based on the camera height and object height. We do not do ori- entation regression in proposal generation, whereas we left it to the next prediction stage. The orientations of 3D boxes
  are restricted to {0◦, 90◦}, which are close to the actual ori- entations of most road scene objects. This simplification makes training of proposal regression easier.

关于该部分的实现待理解之后进行分析。此外，作者考虑到该卷积网络对精度较低的物体的检测效率不高，因此使用到了上采样。

后面的3D Regression的过程真的就是2D上的升维，从实现上来看，基本上没有变化，熟悉2D bounding-box regression的，基本很快就能理解。详情可参看论文。

在利用鸟瞰图生成3D proposal的时候，使用到的就是NMS方法。减少冗余计算。