Yolov5 input (II) | CSDN creative punch in

Beginner little rookie , I hope it's like taking notes and recording what I've learned , Also hope to help the same entry-level people , I hope the big guys can help correct it ~ Tort made delete .

Catalog

One 、 Adaptive anchor frame calculation

Two 、 Adaptive image scaling

1、 Principle analysis

2、 The code analysis

---

One 、 Adaptive anchor frame calculation

YOLOv5 For different data sets , There are anchor boxes with initial length and width . In the process of training , Output the prediction box based on the initial anchor box , Then with the real box groundtruth compare , That is, the difference between the real border position and the preset border is calculated , Then reverse update , Iterative network parameters , Keep training .

Anchor Box The definition of ： Described by the height and width of the border , At first glance, you will feel this Anchor Box It's not fixed , It can form countless on the picture . Here we need a central point , And this central point is extracted by the subsequent network Feature Map The point of , So an initial Anchor box There is no need to specify a central location .

stay yolov5s.yaml The initialization in the file is ：

anchors:
  - [10,13, 16,30, 33,23]  # P3/8
  - [30,61, 62,45, 59,119]  # P4/16
  - [116,90, 156,198, 373,326]  # P5/32

stay YOLOv5 This function is optional , If you think it doesn't work well, you can turn it off , Add

“–noautoanchor” Options

The specific choice is train.py In file

parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')

---

Two 、 Adaptive image scaling

1、 Principle analysis

Usually we find pictures of different sizes , But when it comes to online training , It is necessary to ensure that the image size is consistent .

But if we simply use resize, It will cause the distortion of the picture , And then affect our results .

So take a better approach ——letterbox Adaptive image scaling technology

notes ：train It is not used in the process letterbox Adaptive image scaling technology , Only in detect In-process use .

train The same size of the picture is kept because it puts 4 The parts of the picture form a large picture with the same size , So there's no need to use letterbox

See previous blogs for details YOLOv5 Input end （ One ）—— Mosaic Data to enhance |CSDN Creative punch in _tt Ya's blog -CSDN Blog

letterbox Adaptive image scaling technology tries to maintain the aspect ratio , The missing parts shall be filled with gray edges to reach the fixed size .

Next, let's combine it with the code to see its principle

2、 The code analysis

This part is in utils/augmentations.py In the document letterbox function

First, get the height and width of the current picture , Then ensure that the height and width of the transformed image is an integer

def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
    # Resize and pad image while meeting stride-multiple constraints
    shape = im.shape[:2]  # current shape [height, width]
    # Get the height and width of the current picture 
    if isinstance(new_shape, int):# Determine whether the final image height and width is an integer 
        new_shape = (new_shape, new_shape)
    # Save the final picture height and width in the form of pixel segments to new_shape in 

Calculate scale , Take the smallest proportion , And we only shrink, not enlarge

    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])# Calculate scale 
    if not scaleup:  # only scale down, do not scale up (for better val mAP)
        r = min(r, 1.0)# The operation here is to scale down only , Don't zoom in 

Then calculate the length and width of the fill ：

First, calculate the height and width of the scaled image

Then calculate the length to be filled , And again stride Divide by the remainder

among stride Represents the number of samples under the model 2 To the power of ,（ Feel the problem of wild ） stay YOLOV5 The number of middle and lower samples is 5, So for 32

Finally, the filling length is halved , Because we have to divide up and down , left and right

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))# Calculate the length and width of the contracted picture 
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    # Calculate the length and width to be filled 
    if auto:  # minimum rectangle
        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
        #stride Represents the number of samples under the model 2 To the power of ,（ Feel the problem of wild ） stay YOLOV5 The number of middle and lower samples is 5, So for 32
    elif scaleFill:  # stretch
        dw, dh = 0.0, 0.0
        new_unpad = (new_shape[1], new_shape[0])
        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios

    dw /= 2  # divide padding into 2 sides
    dh /= 2
    # And a half （ Up and down , about ）

Transform and return ：

call resize Function to deform , Then finally determine the number to be filled up, down, left and right （ The guarantee is greater than or equal to 0 The integer of ）

Call again copyMakeBorder fill

Finally, return the image after the operation , Scale and fill height and width

    if shape[::-1] != new_unpad:  # resize
        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
        # Deformation operation 
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    # Directly let less than 1 For the 0, Calculate how much to fill in up, down, left, right （ Integers ）
    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    # fill 
    return im, ratio, (dw, dh)

---

You are welcome to criticize and correct in the comment area , Thank you. ~