Interpretation of orientedrcnn papers

  Thesis link

1、 starting point

  First Rotate RPN Scale the dimensions by scattering at various angles anchor Generate OP(Oriented Proposal Abbreviation ), The disadvantages are obvious ： Large amount of computation and memory .
  therefore RoI Transformer Put forward the help level anchor To generate OP, Due to the abandonment of rotation anchor So the calculation goes on . The general process is ：Horizental_Anchor --> RPN --> Horizental Proposal -->Roi Align–> FC --> OP. shortcoming ： Introduced FC and RoIAlign, Leading to the Internet heavy And complicated .
  Therefore, this paper considers direct Horizental_Anchor --> OP, Thus, the model is simplified to reduce the risk of over fitting .

2、 Model structure

  The overall structure is relatively simple , Second order network structure . Let me first introduce Oriented RPN.

2.1. Oriented RPN

  Here, we will focus on the dimension changes of the lower channel ： First 256 Is the feature graph dimension , After that, we sprinkled... At each position of the feature map A(A=3) individual Horizental anchor, after 1*1 The convolution operation of turns the channel dimension into 6*A = 18. each anchor The predicted output value of is 6 Parameters ：(x,y,w,h, delta_alpha, delta_belta), Each represents the center , The width and height of the circumscribed rectangle of the slanted box and the two quantities offset from the center position . The following is an example ：

  In essence OP The expression becomes 6 Parameter representation . It is easy to get by the following formula OP Of v1-v4 The horizontal and vertical coordinates of the four corners .

  As predicted (delta_alpha, delta_belta) It can't be exactly the same , Therefore, the OP It's a parallelogram ： Here's the picture ：

  and RRoI Align Only the features of the slanted rectangle can be extracted , It can not extract inclined parallelogram feature , So this article is in RPN Output OP Another step of decoding process has been added , Change a parallelogram into (x,y,w,h,theta) Format , Convenient for follow-up RRoI Align The extracted features .

  The idea is to calculate the diagonal length of parallelogram first , Short edge resize Just as long as the long side . Post the code here ：

    polys = torch.stack([ga, y1, x2, gb, _ga, y2, x1, _gb], dim=-1) #  Parallelogram four coordinates 
    center = torch.stack([gx, gy, gx, gy, gx, gy, gx, gy], dim=-1)  #  Stack quartic center coordinates 
    center_polys = polys - center     #  Calculate the relative distance between each coordinate and the center 
    diag_len = torch.sqrt(            #  Calculate the diagonal length , Namely   Pythagorean theorem 
        torch.square(center_polys[..., 0::2]) + torch.square(center_polys[..., 1::2]))
    max_diag_len, _ = torch.max(diag_len, dim=-1, keepdim=True)      #  Get the length of the long side 
    diag_scale_factor = max_diag_len / diag_len                      #  get resize The proportion 
    #  Compare the relative distance resize
    center_polys = center_polys * diag_scale_factor.repeat_interleave(2, dim=-1) 
    rectpolys = center_polys + center  #  The relative distance is + The upper center gets four coordinates of the inclined rectangle 
    obboxes = rectpoly2obb(rectpolys).flatten(-2) #  Turn to five point notation 

2.2. RRoI Align

  This section does not cover , It's extraction Oriented RPN Generated OP Eigenvector of .

summary

  Experimental results SOTA, There's no more post here .