PointRCNN之code学习笔记
作者:互联网
input -> rpn -> rpn_cls,rpn_reg, backbone_xyz, backbone_features
rpn_cls, rpn_reg, backbone_xyz -> proposal_layer -> rois, roi_scores
rpn_score_norm = sigmoid(rpn_cls)
seg_mask = rpn_score_norm > score_thresh
pts_depth = norm2(backbone_xyz.z)
rcnn_input = ‘rpn_xyz(backbone_xyz), rpn_features(backbone_features), seg_mask,
roi_boxes3d(rois), pts_depth’
rcnn_int -> rcnn_net -> rcnn_cls, rcnn_reg
如何产生proposals:
1、以每一帧上前景点(总数为N个)为中心,在每个点上,利用回归值以及设置的平均尺寸,生成初始proposals(大小为(batch_size*N, 7), [x,y,z,h,w,l,ry])
2、根据分类得到的得分,进行排序
3、对每一帧上排序后的proposals,根据其坐标z值来查找proposals:
0<z<=40: 取前6300个proposal, 然后将这些proposals投影到BEV,利用NMS(根据阈值设置),找前358个(不足时,保持NMS处理后的个数)
40<z<=80:取前2700个(不足2700时,取其本来有的个数),然后利用NMS, 找前154个(同上)
4、返回生成的bbox3d及其对应的scores.
标签:code,xyz,backbone,proposals,笔记,rcnn,cls,PointRCNN,rpn 来源: https://blog.csdn.net/daideyun/article/details/100127498