本章节说明pointpillars网络后处理使用方式，算子的详细规格信息，请参考《支持Caffe&TensorFlow&ONNX&PyTorch算子清单》>支持Caffe算子清单中的PointPillarDecodeBox~PointPillarDetection算子，三个算子分别进行网络输出的框、分类置信度归一化、最终检测结果的计算。

注意：caffe.proto的LayerParameter中已经添加了对应的参数层信息，用户只需要根据模型的具体参数规格修改网络模型prototxt中的参数。

在LayerParameter中添加如下参数层：

message LayerParameter {
...
  optional PointpillarDecodeBoxParameter decode_box_param=251;
  optional PointpillarClassCalParameter cls_cal_param=1518;
  optional PointpillarPillarDetectionParameter pillar_detection_param=253;
...
}

在caffe.proto中增加如下参数：

message PointpillarClassCalParameter {
   optional bool use_sigmoid = 1 [default = true];
}

message PointpillarPillarDetectionParameter {
   required float thresh = 1;
   optional int32 per_class_num = 2;
}

message PointpillarDecodeBoxParameter {
   repeated float box1_size = 1;
   repeated float box2_size = 2;
   repeated float box3_size = 3;
   repeated float box4_size = 4;
   repeated float rotations = 5;
   optional float x_start = 6 [default = 1];
   optional float y_start = 7 [default = 1];
   optional float x_strides = 8 [default = 1];
   optional float y_strides = 9 [default = 1];
   repeated float z_strides = 10;
}

后处理对应prototxt文件如下：

input: "coor"
input: "class"
input: "head"
input_shape {
  dim: 1
  dim: 56 
  dim: 288
  dim: 288
}
input_shape {
  dim: 1
  dim: 32 
  dim: 288
  dim: 288
}
input_shape {
  dim: 1
  dim: 16
  dim: 288
  dim: 288
}


layer {
  type: 'PointPillarDecodeBox'
  name: 'decode_box'
  bottom: 'coor'
  top: 'decode_coor'
  decode_box_param {
    box1_size: 0.68
    box1_size: 0.67
    box1_size: 1.77
    box2_size: 1.92
    box2_size: 4.74
    box2_size: 1.72
    box3_size: 0.81
    box3_size: 1.99
    box3_size: 1.68
    box4_size: 1.31
    box4_size: 1.21
    box4_size: 1.00
    rotations: 0
    rotations: 1.5708

    x_start: -22.96
    y_start: -22.96
    x_strides: 0.16
    y_strides: 0.16
    z_strides: -0.18
    z_strides: -0.17
    z_strides: -0.28
    z_strides: -0.21
  }
}

layer {
  type: 'ClassCal'
  name: 'cls_cal'
  bottom: 'class'
  bottom: 'head'
  top: 'class_prob'
  top: 'head_direct'
  cls_cal_param {
    use_sigmoid: True 
  }
}

layer {
  type: 'PointPillarDetection'
  name: 'pillar_detection'
  bottom: 'decode_coor'
  bottom: 'class_prob'
  bottom: 'head_direct'
  top: 'box'
  pillar_detection_param {
    thresh: 0.5
  }
}

PointPillars后处理网络模型prototxt修改