YOLOV8改进-添加EIOU,SIOU,AlphaIOU,FocalEIOU,WIOU.

目录

1替换计算框iou的代码

1.1原代码

1.2替换代码

2修改调用函数

2.1原代码

2.2替换代码

3参数修改

---

本文介绍的是在yolov8模型上修改成其他的IOU,各种IOU需要自己测试才知道效果。

1替换计算框iou的代码

        将ultralytics/yolo/utils/metrics.py路径下的def bbox_iou替换为其他代码

1.1原代码

1. def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
2. """
3. Calculate Intersection over Union (IoU) of box1(1, 4) to box2(n, 4).
5. Args:
6. box1 (torch.Tensor): A tensor representing a single bounding box with shape (1, 4).
7. box2 (torch.Tensor): A tensor representing n bounding boxes with shape (n, 4).
8. xywh (bool, optional): If True, input boxes are in (x, y, w, h) format. If False, input boxes are in
9. (x1, y1, x2, y2) format. Defaults to True.
10. GIoU (bool, optional): If True, calculate Generalized IoU. Defaults to False.
11. DIoU (bool, optional): If True, calculate Distance IoU. Defaults to False.
12. CIoU (bool, optional): If True, calculate Complete IoU. Defaults to False.
13. eps (float, optional): A small value to avoid division by zero. Defaults to 1e-7.
15. Returns:
16. (torch.Tensor): IoU, GIoU, DIoU, or CIoU values depending on the specified flags.
17. """
18. # Get the coordinates of bounding boxes
19. if xywh: # transform from xywh to xyxy
20. (x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, -1), box2.chunk(4, -1)
21. w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2
22. b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_
23. b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_
24. else: # x1, y1, x2, y2 = box1
25. b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, -1)
26. b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, -1)
27. w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
28. w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
30. # Intersection area
31. inter = (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp(0) * \
32. (b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1)).clamp(0)
34. # Union Area
35. union = w1 * h1 + w2 * h2 - inter + eps
37. # IoU
38. iou = inter / union
39. if CIoU or DIoU or GIoU:
40. cw = b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) # convex (smallest enclosing box) width
41. ch = b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) # convex height
42. if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
43. c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared
44. rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center dist ** 2
45. if CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
46. v = (4 / math.pi ** 2) * (torch.atan(w2 / h2) - torch.atan(w1 / h1)).pow(2)
47. with torch.no_grad():
48. alpha = v / (v - iou + (1 + eps))
49. return iou - (rho2 / c2 + v * alpha) # CIoU
50. return iou - rho2 / c2 # DIoU
51. c_area = cw * ch + eps # convex area
52. return iou - (c_area - union) / c_area # GIoU https://arxiv.org/pdf/1902.09630.pdf
53. return iou # IoU

1.2替换代码

1. class WIoU_Scale:
2. ''' monotonous: {
3. None: origin v1
4. True: monotonic FM v2
5. False: non-monotonic FM v3
6. }
7. momentum: The momentum of running mean'''
9. iou_mean = 1.
10. monotonous = False
11. _momentum = 1 - 0.5 ** (1 / 7000)
12. _is_train = True
14. def __init__(self, iou):
15. self.iou = iou
16. self._update(self)
18. @classmethod
19. def _update(cls, self):
20. if cls._is_train: cls.iou_mean = (1 - cls._momentum) * cls.iou_mean + \
21. cls._momentum * self.iou.detach().mean().item()
23. @classmethod
24. def _scaled_loss(cls, self, gamma=1.9, delta=3):
25. if isinstance(self.monotonous, bool):
26. if self.monotonous:
27. return (self.iou.detach() / self.iou_mean).sqrt()
28. else:
29. beta = self.iou.detach() / self.iou_mean
30. alpha = delta * torch.pow(gamma, beta - delta)
31. return beta / alpha
32. return 1
34. def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, SIoU=False, EIoU=False, WIoU=False, Focal=False,
35. alpha=1, gamma=0.5, scale=False, eps=1e-7):
36. # Returns Intersection over Union (IoU) of box1(1,4) to box2(n,4)
38. # Get the coordinates of bounding boxes
39. if xywh: # transform from xywh to xyxy
40. (x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, -1), box2.chunk(4, -1)
41. w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2
42. b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_
43. b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_
44. else: # x1, y1, x2, y2 = box1
45. b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, -1)
46. b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, -1)
47. w1, h1 = b1_x2 - b1_x1, (b1_y2 - b1_y1).clamp(eps)
48. w2, h2 = b2_x2 - b2_x1, (b2_y2 - b2_y1).clamp(eps)
50. # Intersection area
51. inter = (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp(0) * \
52. (b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1)).clamp(0)
54. # Union Area
55. union = w1 * h1 + w2 * h2 - inter + eps
56. if scale:
57. self = WIoU_Scale(1 - (inter / union))
59. # IoU
60. # iou = inter / union # ori iou
61. iou = torch.pow(inter / (union + eps), alpha) # alpha iou
62. if CIoU or DIoU or GIoU or EIoU or SIoU or WIoU:
63. cw = b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) # convex (smallest enclosing box) width
64. ch = b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) # convex height
65. if CIoU or DIoU or EIoU or SIoU or WIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
66. c2 = (cw ** 2 + ch ** 2) ** alpha + eps # convex diagonal squared
67. rho2 = (((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + (
68. b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4) ** alpha # center dist ** 2
69. if CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
70. v = (4 / math.pi ** 2) * (torch.atan(w2 / h2) - torch.atan(w1 / h1)).pow(2)
71. with torch.no_grad():
72. alpha_ciou = v / (v - iou + (1 + eps))
73. if Focal:
74. return iou - (rho2 / c2 + torch.pow(v * alpha_ciou + eps, alpha)), torch.pow(inter / (union + eps),
75. gamma) # Focal_CIoU
76. else:
77. return iou - (rho2 / c2 + torch.pow(v * alpha_ciou + eps, alpha)) # CIoU
78. elif EIoU:
79. rho_w2 = ((b2_x2 - b2_x1) - (b1_x2 - b1_x1)) ** 2
80. rho_h2 = ((b2_y2 - b2_y1) - (b1_y2 - b1_y1)) ** 2
81. cw2 = torch.pow(cw ** 2 + eps, alpha)
82. ch2 = torch.pow(ch ** 2 + eps, alpha)
83. if Focal:
84. return iou - (rho2 / c2 + rho_w2 / cw2 + rho_h2 / ch2), torch.pow(inter / (union + eps),
85. gamma) # Focal_EIou
86. else:
87. return iou - (rho2 / c2 + rho_w2 / cw2 + rho_h2 / ch2) # EIou
88. elif SIoU:
89. # SIoU Loss https://arxiv.org/pdf/2205.12740.pdf
90. s_cw = (b2_x1 + b2_x2 - b1_x1 - b1_x2) * 0.5 + eps
91. s_ch = (b2_y1 + b2_y2 - b1_y1 - b1_y2) * 0.5 + eps
92. sigma = torch.pow(s_cw ** 2 + s_ch ** 2, 0.5)
93. sin_alpha_1 = torch.abs(s_cw) / sigma
94. sin_alpha_2 = torch.abs(s_ch) / sigma
95. threshold = pow(2, 0.5) / 2
96. sin_alpha = torch.where(sin_alpha_1 > threshold, sin_alpha_2, sin_alpha_1)
97. angle_cost = torch.cos(torch.arcsin(sin_alpha) * 2 - math.pi / 2)
98. rho_x = (s_cw / cw) ** 2
99. rho_y = (s_ch / ch) ** 2
100. gamma = angle_cost - 2
101. distance_cost = 2 - torch.exp(gamma * rho_x) - torch.exp(gamma * rho_y)
102. omiga_w = torch.abs(w1 - w2) / torch.max(w1, w2)
103. omiga_h = torch.abs(h1 - h2) / torch.max(h1, h2)
104. shape_cost = torch.pow(1 - torch.exp(-1 * omiga_w), 4) + torch.pow(1 - torch.exp(-1 * omiga_h), 4)
105. if Focal:
106. return iou - torch.pow(0.5 * (distance_cost + shape_cost) + eps, alpha), torch.pow(
107. inter / (union + eps), gamma) # Focal_SIou
108. else:
109. return iou - torch.pow(0.5 * (distance_cost + shape_cost) + eps, alpha) # SIou
110. elif WIoU:
111. if Focal:
112. raise RuntimeError("WIoU do not support Focal.")
113. elif scale:
114. return getattr(WIoU_Scale, '_scaled_loss')(self), (1 - iou) * torch.exp(
115. (rho2 / c2)), iou # WIoU https://arxiv.org/abs/2301.10051
116. else:
117. return iou, torch.exp((rho2 / c2)) # WIoU v1
118. if Focal:
119. return iou - rho2 / c2, torch.pow(inter / (union + eps), gamma) # Focal_DIoU
120. else:
121. return iou - rho2 / c2 # DIoU
122. c_area = cw * ch + eps # convex area
123. if Focal:
124. return iou - torch.pow((c_area - union) / c_area + eps, alpha), torch.pow(inter / (union + eps),
125. gamma) # Focal_GIoU https://arxiv.org/pdf/1902.09630.pdf
126. else:
127. return iou - torch.pow((c_area - union) / c_area + eps, alpha) # GIoU https://arxiv.org/pdf/1902.09630.pdf
128. if Focal:
129. return iou, torch.pow(inter / (union + eps), gamma) # Focal_IoU
130. else:
131. return iou # IoU

2修改调用函数

替换ultralytics/yolo/utils/loss.py路径下class BboxLoss类的def forward函数

2.1原代码

1. iou = bbox_iou(pred_bboxes[fg_mask], target_bboxes[fg_mask], xywh=False, CIoU=True)
2. loss_iou = ((1.0 - iou) * weight).sum() / target_scores_sum

2.2替换代码

1. iou = bbox_iou(pred_bboxes[fg_mask], target_bboxes[fg_mask], xywh=False, WIoU=True,scale=True)
2. if type(iou) is tuple:
3. if len(iou) == 2:
4. loss_iou = ((1.0 - iou[0]) * iou[1].detach() * weight).sum() / target_scores_sum
5. else:
6. loss_iou = (iou[0] * iou[1] * weight).sum() / target_scores_sum
7. else:
8. loss_iou = ((1.0 - iou) * weight).sum() / target_scores_sum

3参数修改

将其中的想使用的IOU设置为True，如WIoU=True，保留一个使用的即可。

iou = bbox_iou(pred_bboxes[fg_mask], target_bboxes[fg_mask], xywh=False, WIoU=True,scale=True)

如果出现报错可能是因为其他的类也调用IOU计算，发生冲突。找到其他调用，将其他的修改。

在ultralytics/yolo/utils/tal.py路径下class TaskAlignedAssigner，其中def get_box_metrics函数

将代码

overlaps[mask_gt] = bbox_iou(gt_boxes, pd_boxes, xywh=False, CIoU=True).squeeze(-1).clamp(0)

替换为

overlaps[mask_gt] = bbox_iou(gt_boxes, pd_boxes, xywh=False).squeeze(-1).clamp(0)