Spaces:
Runtime error
Runtime error
init repo
Browse files- README.md +3 -3
- app.py +58 -0
- constants.py +25 -0
- examples/1.jpg +0 -0
- examples/2.jpg +0 -0
- examples/3.jpg +0 -0
- model.pth +3 -0
- model.py +80 -0
- requirements.txt +4 -0
- utils.py +259 -0
README.md
CHANGED
@@ -1,8 +1,8 @@
|
|
1 |
---
|
2 |
title: Stamp Detection
|
3 |
-
emoji:
|
4 |
-
colorFrom:
|
5 |
-
colorTo:
|
6 |
sdk: gradio
|
7 |
sdk_version: 3.35.2
|
8 |
app_file: app.py
|
|
|
1 |
---
|
2 |
title: Stamp Detection
|
3 |
+
emoji: π
|
4 |
+
colorFrom: yellow
|
5 |
+
colorTo: gray
|
6 |
sdk: gradio
|
7 |
sdk_version: 3.35.2
|
8 |
app_file: app.py
|
app.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# AUTOGENERATED! DO NOT EDIT! File to edit: ../app.ipynb.
|
2 |
+
|
3 |
+
# %% auto 0
|
4 |
+
__all__ = ['device', 'model', 'transforms', 'image', 'result', 'examples', 'intf', 'detect_stamps']
|
5 |
+
|
6 |
+
# %% ../app.ipynb 1
|
7 |
+
from model import YOLOStamp
|
8 |
+
from utils import *
|
9 |
+
import torch
|
10 |
+
import gradio as gr
|
11 |
+
import albumentations as A
|
12 |
+
from albumentations.pytorch.transforms import ToTensorV2
|
13 |
+
from PIL import Image
|
14 |
+
|
15 |
+
# %% ../app.ipynb 2
|
16 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
17 |
+
|
18 |
+
model = YOLOStamp()
|
19 |
+
model.load_state_dict(torch.load('model.pth', map_location=torch.device('cpu')))
|
20 |
+
model = model.to(device)
|
21 |
+
model.eval()
|
22 |
+
|
23 |
+
# %% ../app.ipynb 3
|
24 |
+
transforms = A.Compose([
|
25 |
+
A.Resize(height=448, width=448),
|
26 |
+
A.Normalize(),
|
27 |
+
ToTensorV2(p=1.0),
|
28 |
+
])
|
29 |
+
|
30 |
+
# %% ../app.ipynb 7
|
31 |
+
def detect_stamps(image):
|
32 |
+
shape = image.size[:2]
|
33 |
+
image = image.convert('RGB')
|
34 |
+
image = np.array(image)
|
35 |
+
image = transforms(image=image)['image']
|
36 |
+
|
37 |
+
output = model(image.unsqueeze(0).to(device))[0]
|
38 |
+
boxes = output_tensor_to_boxes(output.detach().cpu())
|
39 |
+
boxes = nonmax_suppression(boxes)
|
40 |
+
img = image.permute(1, 2, 0).cpu().numpy()
|
41 |
+
img = visualize_bbox(img.copy(), boxes=boxes, draw_center=False)
|
42 |
+
|
43 |
+
|
44 |
+
img = cv2.resize(img, dsize=shape)
|
45 |
+
|
46 |
+
return Image.fromarray((255. * (img * np.array(STD) + np.array(MEAN))).astype(np.uint8))
|
47 |
+
|
48 |
+
# %% ../app.ipynb 9
|
49 |
+
image = gr.inputs.Image(type="pil")
|
50 |
+
result = gr.outputs.Image(type="pil")
|
51 |
+
examples = ['./examples/1.jpg', './examples/2.jpg', './examples/3.jpg']
|
52 |
+
|
53 |
+
intf = gr.Interface(fn=detect_stamps,
|
54 |
+
inputs=image,
|
55 |
+
outputs=result,
|
56 |
+
title='Stamp detection',
|
57 |
+
examples=examples)
|
58 |
+
intf.launch(inline=False)
|
constants.py
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# shape of input image to YOLO
|
2 |
+
W, H = 448, 448
|
3 |
+
# grid size after last convolutional layer of YOLO
|
4 |
+
S = 7
|
5 |
+
# anchors of YOLO model
|
6 |
+
ANCHORS = [[1.08,1.19],
|
7 |
+
[3.42,4.41],
|
8 |
+
[16.62,10.52]]
|
9 |
+
# number of anchors boxes
|
10 |
+
BOX = len(ANCHORS)
|
11 |
+
# maximum number of stamps on image
|
12 |
+
STAMP_NB_MAX = 10
|
13 |
+
# minimal confidence of presence a stamp in the grid cell
|
14 |
+
OUTPUT_THRESH = 0.76
|
15 |
+
# maximal iou score to consider boxes different
|
16 |
+
IOU_THRESH = 0.3
|
17 |
+
# path to folder containing images
|
18 |
+
IMAGE_FOLDER = './data/images'
|
19 |
+
# path to .cvs file containing annotations
|
20 |
+
ANNOTATIONS_PATH = './data/all_annotations.csv'
|
21 |
+
# standard deviation and mean of pixel values for normalization
|
22 |
+
STD = (0.229, 0.224, 0.225)
|
23 |
+
MEAN = (0.485, 0.456, 0.406)
|
24 |
+
# box color to show the bounding box on image
|
25 |
+
BOX_COLOR = (0, 0, 255)
|
examples/1.jpg
ADDED
examples/2.jpg
ADDED
examples/3.jpg
ADDED
model.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e3e633824bea418bf44bb82f8d5a667b9d7d29b913079c4d8b6ca217f6f383e7
|
3 |
+
size 502816
|
model.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
|
4 |
+
from constants import *
|
5 |
+
|
6 |
+
"""
|
7 |
+
Class for custom activation.
|
8 |
+
"""
|
9 |
+
class SymReLU(nn.Module):
|
10 |
+
def __init__(self, inplace: bool = False):
|
11 |
+
super().__init__()
|
12 |
+
self.inplace = inplace
|
13 |
+
|
14 |
+
def forward(self, input):
|
15 |
+
return torch.min(torch.max(input, -torch.ones_like(input)), torch.ones_like(input))
|
16 |
+
|
17 |
+
def extra_repr(self) -> str:
|
18 |
+
inplace_str = 'inplace=True' if self.inplace else ''
|
19 |
+
return inplace_str
|
20 |
+
|
21 |
+
|
22 |
+
"""
|
23 |
+
Class implementing YOLO-Stamp architecture described in https://link.springer.com/article/10.1134/S1054661822040046.
|
24 |
+
"""
|
25 |
+
class YOLOStamp(nn.Module):
|
26 |
+
def __init__(
|
27 |
+
self,
|
28 |
+
anchors=ANCHORS,
|
29 |
+
in_channels=3,
|
30 |
+
):
|
31 |
+
super().__init__()
|
32 |
+
|
33 |
+
self.register_buffer('anchors', torch.tensor(anchors))
|
34 |
+
|
35 |
+
self.act = SymReLU()
|
36 |
+
self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
|
37 |
+
self.conv1 = nn.Conv2d(in_channels=in_channels, out_channels=8, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
38 |
+
self.norm1 = nn.BatchNorm2d(num_features=8)
|
39 |
+
self.conv2 = nn.Conv2d(in_channels=8, out_channels=16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
40 |
+
self.norm2 = nn.BatchNorm2d(num_features=16)
|
41 |
+
self.conv3 = nn.Conv2d(in_channels=16, out_channels=16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
42 |
+
self.norm3 = nn.BatchNorm2d(num_features=16)
|
43 |
+
self.conv4 = nn.Conv2d(in_channels=16, out_channels=16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
44 |
+
self.norm4 = nn.BatchNorm2d(num_features=16)
|
45 |
+
self.conv5 = nn.Conv2d(in_channels=16, out_channels=16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
46 |
+
self.norm5 = nn.BatchNorm2d(num_features=16)
|
47 |
+
self.conv6 = nn.Conv2d(in_channels=16, out_channels=24, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
48 |
+
self.norm6 = nn.BatchNorm2d(num_features=24)
|
49 |
+
self.conv7 = nn.Conv2d(in_channels=24, out_channels=24, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
50 |
+
self.norm7 = nn.BatchNorm2d(num_features=24)
|
51 |
+
self.conv8 = nn.Conv2d(in_channels=24, out_channels=48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
52 |
+
self.norm8 = nn.BatchNorm2d(num_features=48)
|
53 |
+
self.conv9 = nn.Conv2d(in_channels=48, out_channels=48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
54 |
+
self.norm9 = nn.BatchNorm2d(num_features=48)
|
55 |
+
self.conv10 = nn.Conv2d(in_channels=48, out_channels=48, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
56 |
+
self.norm10 = nn.BatchNorm2d(num_features=48)
|
57 |
+
self.conv11 = nn.Conv2d(in_channels=48, out_channels=64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
|
58 |
+
self.norm11 = nn.BatchNorm2d(num_features=64)
|
59 |
+
self.conv12 = nn.Conv2d(in_channels=64, out_channels=256, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
|
60 |
+
self.norm12 = nn.BatchNorm2d(num_features=256)
|
61 |
+
self.conv13 = nn.Conv2d(in_channels=256, out_channels=len(anchors) * 5, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
|
62 |
+
|
63 |
+
def forward(self, x, head=True):
|
64 |
+
x = x.type(self.conv1.weight.dtype)
|
65 |
+
x = self.act(self.pool(self.norm1(self.conv1(x))))
|
66 |
+
x = self.act(self.pool(self.norm2(self.conv2(x))))
|
67 |
+
x = self.act(self.pool(self.norm3(self.conv3(x))))
|
68 |
+
x = self.act(self.pool(self.norm4(self.conv4(x))))
|
69 |
+
x = self.act(self.pool(self.norm5(self.conv5(x))))
|
70 |
+
x = self.act(self.norm6(self.conv6(x)))
|
71 |
+
x = self.act(self.norm7(self.conv7(x)))
|
72 |
+
x = self.act(self.pool(self.norm8(self.conv8(x))))
|
73 |
+
x = self.act(self.norm9(self.conv9(x)))
|
74 |
+
x = self.act(self.norm10(self.conv10(x)))
|
75 |
+
x = self.act(self.norm11(self.conv11(x)))
|
76 |
+
x = self.act(self.norm12(self.conv12(x)))
|
77 |
+
x = self.conv13(x)
|
78 |
+
nb, _, nh, nw= x.shape
|
79 |
+
x = x.permute(0, 2, 3, 1).view(nb, nh, nw, self.anchors.shape[0], 5)
|
80 |
+
return x
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
opencv-python
|
3 |
+
pillow
|
4 |
+
albumentations
|
utils.py
ADDED
@@ -0,0 +1,259 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import cv2
|
3 |
+
import pandas as pd
|
4 |
+
import numpy as np
|
5 |
+
from pathlib import Path
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
from constants import *
|
8 |
+
|
9 |
+
|
10 |
+
def output_tensor_to_boxes(boxes_tensor):
|
11 |
+
"""
|
12 |
+
Converts the YOLO output tensor to list of boxes with probabilites.
|
13 |
+
|
14 |
+
Arguments:
|
15 |
+
boxes_tensor -- tensor of shape (S, S, BOX, 5)
|
16 |
+
|
17 |
+
Returns:
|
18 |
+
boxes -- list of shape (None, 5)
|
19 |
+
|
20 |
+
Note: "None" is here because you don't know the exact number of selected boxes, as it depends on the threshold.
|
21 |
+
For example, the actual output size of scores would be (10, 5) if there are 10 boxes
|
22 |
+
"""
|
23 |
+
cell_w, cell_h = W/S, H/S
|
24 |
+
boxes = []
|
25 |
+
|
26 |
+
for i in range(S):
|
27 |
+
for j in range(S):
|
28 |
+
for b in range(BOX):
|
29 |
+
anchor_wh = torch.tensor(ANCHORS[b])
|
30 |
+
data = boxes_tensor[i,j,b]
|
31 |
+
xy = torch.sigmoid(data[:2])
|
32 |
+
wh = torch.exp(data[2:4])*anchor_wh
|
33 |
+
obj_prob = torch.sigmoid(data[4])
|
34 |
+
|
35 |
+
if obj_prob > OUTPUT_THRESH:
|
36 |
+
x_center, y_center, w, h = xy[0], xy[1], wh[0], wh[1]
|
37 |
+
x, y = x_center+j-w/2, y_center+i-h/2
|
38 |
+
x,y,w,h = x*cell_w, y*cell_h, w*cell_w, h*cell_h
|
39 |
+
box = [x,y,w,h, obj_prob]
|
40 |
+
boxes.append(box)
|
41 |
+
return boxes
|
42 |
+
|
43 |
+
|
44 |
+
def plot_img(img, size=(7,7)):
|
45 |
+
plt.figure(figsize=size)
|
46 |
+
plt.imshow(img)
|
47 |
+
plt.show()
|
48 |
+
|
49 |
+
|
50 |
+
def plot_normalized_img(img, std=STD, mean=MEAN, size=(7,7)):
|
51 |
+
mean = mean if isinstance(mean, np.ndarray) else np.array(mean)
|
52 |
+
std = std if isinstance(std, np.ndarray) else np.array(std)
|
53 |
+
plt.figure(figsize=size)
|
54 |
+
plt.imshow((255. * (img * std + mean)).astype(np.uint))
|
55 |
+
plt.show()
|
56 |
+
|
57 |
+
|
58 |
+
def visualize_bbox(img, boxes, thickness=2, color=BOX_COLOR, draw_center=True):
|
59 |
+
"""
|
60 |
+
Draws boxes on the given image.
|
61 |
+
|
62 |
+
Arguments:
|
63 |
+
img -- torch.Tensor of shape (3, W, H) or numpy.ndarray of shape (W, H, 3)
|
64 |
+
boxes -- list of shape (None, 5)
|
65 |
+
thickness -- number specifying the thickness of box border
|
66 |
+
color -- RGB tuple of shape (3,) specifying the color of boxes
|
67 |
+
draw_center -- boolean specifying whether to draw center or not
|
68 |
+
|
69 |
+
Returns:
|
70 |
+
img_copy -- numpy.ndarray of shape(W, H, 3) containing image with bouning boxes
|
71 |
+
"""
|
72 |
+
img_copy = img.cpu().permute(1,2,0).numpy() if isinstance(img, torch.Tensor) else img.copy()
|
73 |
+
for box in boxes:
|
74 |
+
x,y,w,h = int(box[0]), int(box[1]), int(box[2]), int(box[3])
|
75 |
+
img_copy = cv2.rectangle(
|
76 |
+
img_copy,
|
77 |
+
(x,y),(x+w, y+h),
|
78 |
+
color, thickness)
|
79 |
+
if draw_center:
|
80 |
+
center = (x+w//2, y+h//2)
|
81 |
+
img_copy = cv2.circle(img_copy, center=center, radius=3, color=(0,255,0), thickness=2)
|
82 |
+
return img_copy
|
83 |
+
|
84 |
+
|
85 |
+
def read_data(annotations=Path(ANNOTATIONS_PATH)):
|
86 |
+
"""
|
87 |
+
Reads annotations data from .csv file. Must contain columns: image_name, bbox_x, bbox_y, bbox_width, bbox_height.
|
88 |
+
|
89 |
+
Arguments:
|
90 |
+
annotations_path -- string or Path specifying path of annotations file
|
91 |
+
|
92 |
+
Returns:
|
93 |
+
data -- list of dictionaries containing path, number of boxes and boxes itself
|
94 |
+
"""
|
95 |
+
data = []
|
96 |
+
|
97 |
+
boxes = pd.read_csv(annotations)
|
98 |
+
image_names = boxes['image_name'].unique()
|
99 |
+
|
100 |
+
for image_name in image_names:
|
101 |
+
cur_boxes = boxes[boxes['image_name'] == image_name]
|
102 |
+
img_data = {
|
103 |
+
'file_path': image_name,
|
104 |
+
'box_nb': len(cur_boxes),
|
105 |
+
'boxes': []}
|
106 |
+
stamp_nb = img_data['box_nb']
|
107 |
+
if stamp_nb <= STAMP_NB_MAX:
|
108 |
+
img_data['boxes'] = cur_boxes[['bbox_x', 'bbox_y','bbox_width','bbox_height']].values
|
109 |
+
data.append(img_data)
|
110 |
+
return data
|
111 |
+
|
112 |
+
|
113 |
+
def boxes_to_tensor(boxes):
|
114 |
+
"""
|
115 |
+
Convert list of boxes (and labels) to tensor format
|
116 |
+
|
117 |
+
Arguments:
|
118 |
+
boxes -- list of boxes
|
119 |
+
|
120 |
+
Returns:
|
121 |
+
boxes_tensor -- tensor of shape (S, S, BOX, 5)
|
122 |
+
"""
|
123 |
+
boxes_tensor = torch.zeros((S, S, BOX, 5))
|
124 |
+
cell_w, cell_h = W/S, H/S
|
125 |
+
for i, box in enumerate(boxes):
|
126 |
+
x, y, w, h = box
|
127 |
+
# normalize xywh with cell_size
|
128 |
+
x, y, w, h = x / cell_w, y / cell_h, w / cell_w, h / cell_h
|
129 |
+
center_x, center_y = x + w / 2, y + h / 2
|
130 |
+
grid_x = int(np.floor(center_x))
|
131 |
+
grid_y = int(np.floor(center_y))
|
132 |
+
|
133 |
+
if grid_x < S and grid_y < S:
|
134 |
+
boxes_tensor[grid_y, grid_x, :, 0:4] = torch.tensor(BOX * [[center_x - grid_x, center_y - grid_y, w, h]])
|
135 |
+
boxes_tensor[grid_y, grid_x, :, 4] = torch.tensor(BOX * [1.])
|
136 |
+
return boxes_tensor
|
137 |
+
|
138 |
+
|
139 |
+
def target_tensor_to_boxes(boxes_tensor, output_threshold=OUTPUT_THRESH):
|
140 |
+
"""
|
141 |
+
Recover target tensor (tensor output of dataset) to bboxes.
|
142 |
+
Arguments:
|
143 |
+
boxes_tensor -- tensor of shape (S, S, BOX, 5)
|
144 |
+
Returns:
|
145 |
+
boxes -- list of boxes, each box is [x, y, w, h]
|
146 |
+
"""
|
147 |
+
cell_w, cell_h = W/S, H/S
|
148 |
+
boxes = []
|
149 |
+
for i in range(S):
|
150 |
+
for j in range(S):
|
151 |
+
for b in range(BOX):
|
152 |
+
data = boxes_tensor[i,j,b]
|
153 |
+
x_center,y_center, w, h, obj_prob = data[0], data[1], data[2], data[3], data[4]
|
154 |
+
if obj_prob > output_threshold:
|
155 |
+
x, y = x_center+j-w/2, y_center+i-h/2
|
156 |
+
x,y,w,h = x*cell_w, y*cell_h, w*cell_w, h*cell_h
|
157 |
+
box = [x,y,w,h]
|
158 |
+
boxes.append(box)
|
159 |
+
return boxes
|
160 |
+
|
161 |
+
|
162 |
+
def overlap(interval_1, interval_2):
|
163 |
+
"""
|
164 |
+
Calculates length of overlap between two intervals.
|
165 |
+
|
166 |
+
Arguments:
|
167 |
+
interval_1 -- list or tuple of shape (2,) containing endpoints of the first interval
|
168 |
+
interval_2 -- list or tuple of shape (2, 2) containing endpoints of the second interval
|
169 |
+
|
170 |
+
Returns:
|
171 |
+
overlap -- length of overlap
|
172 |
+
"""
|
173 |
+
x1, x2 = interval_1
|
174 |
+
x3, x4 = interval_2
|
175 |
+
if x3 < x1:
|
176 |
+
if x4 < x1:
|
177 |
+
return 0
|
178 |
+
else:
|
179 |
+
return min(x2,x4) - x1
|
180 |
+
else:
|
181 |
+
if x2 < x3:
|
182 |
+
return 0
|
183 |
+
else:
|
184 |
+
return min(x2,x4) - x3
|
185 |
+
|
186 |
+
|
187 |
+
def compute_iou(box1, box2):
|
188 |
+
"""
|
189 |
+
Compute IOU between box1 and box2.
|
190 |
+
|
191 |
+
Argmunets:
|
192 |
+
box1 -- list of shape (5, ). Represents the first box
|
193 |
+
box2 -- list of shape (5, ). Represents the second box
|
194 |
+
Each box is [x, y, w, h, prob]
|
195 |
+
|
196 |
+
Returns:
|
197 |
+
iou -- intersection over union score between two boxes
|
198 |
+
"""
|
199 |
+
x1,y1,w1,h1 = box1[0], box1[1], box1[2], box1[3]
|
200 |
+
x2,y2,w2,h2 = box2[0], box2[1], box2[2], box2[3]
|
201 |
+
|
202 |
+
area1, area2 = w1*h1, w2*h2
|
203 |
+
intersect_w = overlap((x1,x1+w1), (x2,x2+w2))
|
204 |
+
intersect_h = overlap((y1,y1+h1), (y2,y2+w2))
|
205 |
+
if intersect_w == w1 and intersect_h == h1 or intersect_w == w2 and intersect_h == h2:
|
206 |
+
return 1.
|
207 |
+
intersect_area = intersect_w*intersect_h
|
208 |
+
iou = intersect_area/(area1 + area2 - intersect_area)
|
209 |
+
return iou
|
210 |
+
|
211 |
+
|
212 |
+
def nonmax_suppression(boxes, iou_thresh = IOU_THRESH):
|
213 |
+
"""
|
214 |
+
Removes ovelap bboxes
|
215 |
+
|
216 |
+
Arguments:
|
217 |
+
boxes -- list of shape (None, 5)
|
218 |
+
iou_thresh -- maximal value of iou when boxes are considered different
|
219 |
+
Each box is [x, y, w, h, prob]
|
220 |
+
|
221 |
+
Returns:
|
222 |
+
boxes -- list of shape (None, 5) with removed overlapping boxes
|
223 |
+
"""
|
224 |
+
boxes = sorted(boxes, key=lambda x: x[4], reverse=True)
|
225 |
+
for i, current_box in enumerate(boxes):
|
226 |
+
if current_box[4] <= 0:
|
227 |
+
continue
|
228 |
+
for j in range(i+1, len(boxes)):
|
229 |
+
iou = compute_iou(current_box, boxes[j])
|
230 |
+
if iou > iou_thresh:
|
231 |
+
boxes[j][4] = 0
|
232 |
+
boxes = [box for box in boxes if box[4] > 0]
|
233 |
+
return boxes
|
234 |
+
|
235 |
+
|
236 |
+
|
237 |
+
def yolo_head(yolo_output):
|
238 |
+
"""
|
239 |
+
Converts a yolo output tensor to separate tensors of coordinates, shapes and probabilities.
|
240 |
+
|
241 |
+
Arguments:
|
242 |
+
yolo_output -- tensor of shape (batch_size, S, S, BOX, 5)
|
243 |
+
|
244 |
+
Returns:
|
245 |
+
xy -- tensor of shape (batch_size, S, S, BOX, 2) containing coordinates of centers of found boxes for each anchor in each grid cell
|
246 |
+
wh -- tensor of shape (batch_size, S, S, BOX, 2) containing width and height of found boxes for each anchor in each grid cell
|
247 |
+
prob -- tensor of shape (batch_size, S, S, BOX, 1) containing the probability of presence of boxes for each anchor in each grid cell
|
248 |
+
"""
|
249 |
+
xy = torch.sigmoid(yolo_output[..., 0:2])
|
250 |
+
anchors_wh = torch.tensor(ANCHORS, device=yolo_output.device).view(1, 1, 1, len(ANCHORS), 2)
|
251 |
+
wh = torch.exp(yolo_output[..., 2:4]) * anchors_wh
|
252 |
+
prob = torch.sigmoid(yolo_output[..., 4:5])
|
253 |
+
return xy, wh, prob
|
254 |
+
|
255 |
+
def process_target(target):
|
256 |
+
xy = target[..., 0:2]
|
257 |
+
wh = target[..., 2:4]
|
258 |
+
prob = target[..., 4:5]
|
259 |
+
return xy, wh, prob
|