本章节说明DFMBPSROIAlign、DetectionOutputSSD、BBoxReg三个算子的prototxt使用，很多参数都依据开源网络设置，均为固定值，算子详细参数规格请参考《支持Caffe&TensorFlow&ONNX算子清单》>支持Caffe算子清单中的DFMBPSROIAlign~RPNProposalSSD算子，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 DFMBPSROIAlignParameter dfmb_psroi_pooling_param = 231;
 optional DetectionOutputSSDParameter detection_output_ssd_param = 1510;
 optional BBoxRegParameter bbox_reg_param = 1511;
...
}

在caffe.proto中增加如下参数：

message DFMBPSROIAlignParameter {
   optional uint32 class_num = 1 [default = 10];
   optional uint32 group_height = 2 [default = 7]; 
   optional uint32 group_width = 3 [default = 7]; 
   optional uint32 pooled_height = 4 [default = 7]; 
   optional uint32 pooled_width = 5 [default = 7]; 
   optional uint32 sample_per_part = 6 [default = 4];
   optional float heat_map_a = 7;
}

message DetectionOutputSSDParameter{
  enum MIN_SIZE_MODE {
    HEIGHT_AND_WIDTH = 0;
    HEIGHT_OR_WIDTH = 1;
  }
  optional NMSSSDParameter nms_param = 1;
  repeated float        threshold = 2;
  optional uint32       channel_per_scale = 3 [default = 5];
  optional string       class_name_list = 4 ;
  optional uint32       num_class = 5 [default = 1];
  optional bool         refine_out_of_map_bbox = 6 [default = false];
  repeated uint32       class_indexes = 7;
  repeated float        heat_map_a = 8;
  repeated float        heat_map_b = 13;
  optional float        threshold_objectness = 9 [default = 0.0];
  repeated float        proposal_min_sqrt_area = 10;
  repeated float        proposal_max_sqrt_area = 11;
  optional bool         bg_as_one_of_softmax = 12 [default = false];
  optional bool use_target_type_rcnn = 14 [default = true];
  optional float im_width = 15 [default = 0];
  optional float im_height = 16 [default = 0];
  optional bool rpn_proposal_output_score = 17 [default = false];
  optional bool regress_agnostic = 18 [default = true];
  optional GenerateAnchorParameter gen_anchor_param = 19;
  optional float allow_border = 20 [default = -1.0];
  optional float allow_border_ratio = 21 [default = -1.0];
  optional bool bbox_size_add_one = 22 [default = true];
  optional float read_width_scale = 29 [default = 1.0];
  optional float read_height_scale = 30 [default = 1.0];
  optional uint32 read_height_offset = 31 [default = 0];
  optional float min_size_h = 32 [default = 2.0];
  optional float min_size_w = 33 [default = 2.0];
  optional MIN_SIZE_MODE min_size_mode = 34 [default = HEIGHT_AND_WIDTH];
  optional KPTSParameter kpts_param = 35;
  optional ATRSParameter atrs_param = 36;
  optional FTRSParameter ftrs_param = 37;
  optional SPMPParameter spmp_param = 38;
  optional Cam3dParameter cam3d_param = 39;
}

message BBoxRegParameter {
  repeated float bbox_mean = 1;
  repeated float bbox_std = 2;
}
message NMSSSDParameter{
  optional bool     need_nms = 1 [default = true];
  repeated float    overlap_ratio = 2 ;
  repeated uint32   top_n = 3 ;
  optional bool     add_score = 4 [default = false];
  repeated int32    max_candidate_n = 5;
  repeated bool     use_soft_nms = 6;
  optional bool     nms_among_classes = 7 [default = false];
  repeated bool     voting = 8;
  repeated float    vote_iou = 9;
  optional int32 nms_gpu_max_n_per_time = 12 [default = -1];
}
message GenerateAnchorParameter{
  optional float base_size = 1 [default = 16];
  repeated float ratios = 2;
  repeated float scales = 3;
  repeated float anchor_width = 4;
  repeated float anchor_height = 5;
  repeated float anchor_x1 = 6;
  repeated float anchor_y1 = 7;
  repeated float anchor_x2 = 8;
  repeated float anchor_y2 = 9;
  optional bool zero_anchor_center = 10 [default = true];
}
message KPTSParameter{
  required uint32 kpts_exist_bottom_idx = 1;
  required uint32 kpts_reg_bottom_idx = 2;
  optional bool kpts_reg_as_classify = 3 [default = false];
  optional uint32 kpts_classify_width = 4;
  optional uint32 kpts_classify_height = 5;
  optional int32 kpts_reg_norm_idx_st = 6;
  repeated int32 kpts_st_for_each_class = 7;
  repeated int32 kpts_ed_for_each_class = 8;
  optional float kpts_classify_pad_ratio = 9 [default = 0.0];
}

message ATRSParameter{
  enum NormType {
    NONE = 0;
    WIDTH = 1;
    HEIGHT = 2;
    WIDTH_LOG = 3;
    HEIGHT_LOG = 4;
  }
  required uint32 atrs_reg_bottom_idx = 1;
  optional int32 atrs_reg_norm_idx_st = 2;
  repeated NormType atrs_norm_type = 3;
}
message FTRSParameter{
  required uint32 ftrs_bottom_idx = 1;
}
message SPMPParameter{
  required uint32 spmp_bottom_idx = 1;
  repeated bool spmp_class_aware = 2;
  repeated uint32 spmp_label_width = 3;
  repeated uint32 spmp_label_height = 4;
  repeated float spmp_pad_ratio = 5;
}
message Cam3dParameter{
  required uint32 cam3d_bottom_idx = 1;
}

单算子对应的prototxt样例如下：

DFMBPSROIAlign的prototxt参考：

input: "ft_add_left_right"
input: "rois"
input_shape {
  dim: 1
  dim: 490
  dim: 32
  dim: 32
}
input_shape {
  dim: 1
  dim: 300
  dim: 8
  dim: 1
}
layer {
  type: 'DFMBPSROIAlign'
  name: 'psroi_rois'
  bottom: 'ft_add_left_right'
  bottom: 'rois'
  top: 'psroi_rois'
  dfmb_psroi_pooling_param {
    heat_map_a: 8
    output_dim: 10
    group_height: 7
    group_width: 7
    pooled_height: 7
    pooled_width: 7
    pad_ratio: 0.000000
    sample_per_part: 4
  }
}

RPNProposalSSD的prototxt参考：

input: "rpn_cls_prob_reshape"
input_shape {
  dim: 1
  dim: 30
  dim: 32
  dim: 32
}
input: "rpn_bbox_pred"
input_shape {
  dim: 1
  dim: 60
  dim: 32
  dim: 32
}
input: "im_info"
input_shape {
  dim: 1
  dim: 6
  dim: 1
  dim: 1
}
layer {
  type: 'RPNProposalSSD'
  name: 'proposal'
  bottom: 'rpn_cls_prob_reshape'
  bottom: 'rpn_bbox_pred'
  bottom: 'im_info'
  top: 'rois'
  bbox_reg_param {
    bbox_mean: 0.000437
    bbox_mean: 0.002586
    bbox_mean: -0.123953
    bbox_mean: -0.081469
    bbox_std: 0.126770
    bbox_std: 0.095741
    bbox_std: 0.317300
    bbox_std: 0.281042
  }
  detection_output_ssd_param {
    heat_map_a: 8
    min_size_h: 6.160560
    min_size_w: 6.160560
    min_size_mode: HEIGHT_OR_WIDTH
    threshold_objectness: 0.200000
    gen_anchor_param {
      anchor_width: 9.232984
      anchor_height: 27.726680
      anchor_width: 16.000000
      anchor_height: 16.000000
      anchor_width: 27.712813
      anchor_height: 9.237604
      anchor_width: 18.465969
      anchor_height: 55.453359
      anchor_width: 32.000000
      anchor_height: 32.000000
      anchor_width: 55.425626
      anchor_height: 18.475209
      anchor_width: 36.931937
      anchor_height: 110.906719
      anchor_width: 64.000000
      anchor_height: 64.000000
      anchor_width: 110.851252
      anchor_height: 36.950417
      anchor_width: 73.863875
      anchor_height: 221.813438
      anchor_width: 128.000000
      anchor_height: 128.000000
      anchor_width: 221.702503
      anchor_height: 73.900834
      anchor_width: 147.727750
      anchor_height: 443.626876
      anchor_width: 256.000000
      anchor_height: 256.000000
      anchor_width: 443.405007
      anchor_height: 147.801669
    }
    refine_out_of_map_bbox: true
    nms_param {
      overlap_ratio: 0.700000
      top_n: 300
      max_candidate_n: 3000
      use_soft_nms: false
      voting: false
      vote_iou: 0.700000
    }
  }
}

RCNNProposal的prototxt参考样例如下：

input: "cls_score_softmax"
input_shape {
  dim: 300
  dim: 4
  dim: 1
  dim: 1
}

input: "bbox_pred"
input_shape {
  dim: 300
  dim: 16
  dim: 1
  dim: 1
}

input: "rois"
input_shape {
  dim: 1
  dim: 300
  dim: 8
  dim: 1
}

input: "im_info"
input_shape {
  dim: 1
  dim: 6
  dim: 1
  dim: 1
}

layer {
  type: 'RCNNProposal'
  name: 'rcnn_proposal'
  bottom: 'cls_score_softmax'
  bottom: 'bbox_pred'
  bottom: 'rois'
  bottom: 'im_info'
  top: 'bboxes'
  bbox_reg_param {
    bbox_mean: 0.000000
    bbox_mean: 0.000000
    bbox_mean: 0.000000
    bbox_mean: 0.000000
    bbox_std: 0.100000
    bbox_std: 0.100000
    bbox_std: 0.200000
    bbox_std: 0.200000
  }
  detection_output_ssd_param {
    num_class: 3
    rpn_proposal_output_score: true
    regress_agnostic: false
    min_size_h: 8.800800
    min_size_w: 8.800800
    min_size_mode: HEIGHT_OR_WIDTH
    threshold_objectness: 0.100000
    threshold: 0.100000
    threshold: 0.100000
    threshold: 0.100000
    refine_out_of_map_bbox: true
    nms_param {
      overlap_ratio: 0.500000
      top_n: 5
      max_candidate_n: 300
      use_soft_nms: false
      voting: false
      vote_iou: 0.600000
    }
  }
}

traffic_light_detection网络模型prototxt修改