RCNNProposal

Input

• cls_score_softmax:
  • Required: yes
  • Type: float16
  • Description: confidence value
  • Restriction: NCHW, with shape [b * topn1, 4, 1, 1], where b indicates the batch size and topn1 equals top_n of the RPNProposalSSD layer.
• bbox_pred:
  • Required: yes
  • Type: float16
  • Description: coordinate offset of the BBox, which is used to fine-tune rois
  • Restriction: NCHW. Has shape is [b * topn1, 8, 1, 1] (foreground and background) when regress_agnostic attribute is set to true; else, [b* topn1, 16, 1, 1] (background and three classes), where, b indicates the batch size and topn1 equals top_n of the RPNProposalSSD layer.
• rois:
  • Required: yes
  • Type: float16
  • Description: BBox coordinates
  • Restriction: NCHW, with shape [b, topn1, 8, 1], where b indicates the batch size and topn1 equals top_n of the RPNProposalSSD layer.
• im_info:
  • Required: yes
  • Type: float16
  • Description: image information
  • Restriction: NCHW, with shape [b, 6, 1, 1], where b indicates the batch size

Attribute

• bbox_mean:
  • Required: yes
  • Type: listFloat
  • Description: mean value for BBox normalization
  • Restriction: either 0 or 4

• bbox_std:
  • Required: yes
  • Type: listFloat
  • Description: standard deviation for BBox normalization
  • Restriction: either 0 or 4

• num_class:
  • Required: no
  • Type: int
  • Description: number of classes
  • Restriction: Must be 3.

• rpn_proposal_output_score:
  • Required: no
  • Type: bool
  • Description: whether to output the confidence value of each class
  • Restriction: Must be true.

• regress_agnostic:
  • Required: no
  • Type: bool
  • Description: whether to use the foreground class
  • Restriction: none

• min_size_h:
  • Required: yes
  • Type: float
  • Description: minimum height of the BBox (default = 8.8008)
  • Restriction: none

• min_size_w:
  • Required: yes
  • Type: float
  • Description: minimum width of the BBox (default = 8.8008)
  • Restriction: none

• min_size_mode:
  • Required: yes
  • Type: str
  • Description: minimum size mode enum, default: HEIGHT_OR_WIDTH (HEIGHT_OR_WIDTH=0,HEIGHT_OR_WIDTH=1)
  • Restriction: none

• threshold_objectness:
  • Required: yes
  • Type: float
  • Description: foreground threshold
  • Restriction: Must be in the range (0, 1).

• threshold:
  • Required: yes
  • Type: listFloat
  • Description: confidence threshold of each class (size = num_class)
  • Restriction: Must be in the range (0, 1).

• refine_out_of_map_bbox:
  • Required: yes
  • Type: bool
  • Description: whether to refine out-of-map BBoxes
  • Restriction: none

• overlap_ratio:
  • Required: yes
  • Type: listFloat
  • Description: Intersection over Union (IoU) threshold for NMS
  • Restriction: none

• top_n:
  • Required: yes
  • Type: listInt
  • Description: number of BBoxes after the NMS operation
  • Restriction: none

• max_candidate_n:
  • Required: yes
  • Type: listInt
  • Description: maximum number of candidate BBoxes for NMS
  • Restriction: none

• bsz01:
  • Required: no
  • Type: float
  • Description: BBox size scaling factor
  • Restriction: none

• allow_border:
  • Required: no
  • Type: float
  • Description: allowed border, in pixels
  • Restriction: none

• allow_border_ratio:
  • Required: no
  • Type: float
  • Description: allowed border ratio
  • Restriction: none

Output