【YOLO改进】换遍IoU损失函数之Focal IoU Loss（基于MMYOLO）

Focal IoU损失函数

在目标检测任务中，评估预测边界框的质量是一个重要环节。传统的 IoU（Intersection over Union）损失函数虽然能够评估预测边界框与真实边界框的重叠程度，但在某些情况下存在一些问题。例如，当预测边界框与真实边界框的重叠度较低时，IoU 损失函数的梯度会非常小，导致模型难以进行优化。此外，正负样本分布不均衡也是目标检测任务中的一个常见问题，这会影响模型的训练效果。

为了解决这些问题，研究者提出了 Focal IoU Loss。该损失函数旨在平衡高质量样本和低质量样本对损失的贡献，同时强调困难样本在训练过程中的重要性。

设计原理

Focal IoU Loss 的设计原理主要基于以下两个方面：

1. 平衡高质量样本和低质量样本对损失的贡献：通过引入一个调节因子，使得高质量样本（即 IoU 值较大的样本）在损失计算中占据更大的权重，而低质量样本（即 IoU 值较小的样本）的权重则相对较小。这样可以在一定程度上避免低质量样本对模型训练的负面影响。
2. 强调困难样本在训练过程中的重要性：类似于 Focal Loss，Focal IoU Loss 也为具有挑战性的样本（即 IoU 值较低的样本）分配更高的权重。这样可以使模型在训练过程中更加关注这些困难样本，从而提高模型的泛化能力。

计算步骤

Focal IoU Loss 的计算步骤大致如下：

1. 计算预测边界框与真实边界框的 IoU 值。
2. 根据 IoU 值的大小，将样本分为高质量样本和低质量样本。具体划分标准可以根据实际情况进行调整。
3. 引入一个调节因子 α，用于平衡高质量样本和低质量样本对损失的贡献。α 的取值可以根据实际情况进行调整，通常取值在 0 到 1 之间。
4. 对于高质量样本，直接使用原始的 IoU 损失函数进行计算；对于低质量样本，则根据 IoU 值的大小为其分配一个较小的权重。
5. 将所有样本的损失进行加权求和，得到最终的 Focal IoU Loss。

添加 Focal IoU损失函数(基于MMYOLO)

由于MMYOLO中没有实现Focal IoU损失函数，所以需要在mmyolo/models/iou_loss.py中添加Focal IoU的计算和对应的iou_mode，修改完以后在终端运行

python setup.py install

再在配置文件中进行修改即可。

包含前面所有IoU损失函数替换的mmyolo/models/iou_loss.py如下（Focal IoU只实现了Focal CIoU，其他同理）：

# Copyright (c) OpenMMLab. All rights reserved.
import math
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
from mmdet.models.losses.utils import weight_reduce_loss
from mmdet.structures.bbox import HorizontalBoxes

from mmyolo.registry import MODELS
class WIoU_Scale:
    ''' monotonous: {
            None: origin v1
            True: monotonic FM v2
            False: non-monotonic FM v3
        }
        momentum: The momentum of running mean'''

    iou_mean = 1.
    monotonous = False
    _momentum = 1 - 0.5 ** (1 / 7000)
    _is_train = True

    def __init__(self, iou):
        self.iou = iou
        self._update(self)

    @classmethod
    def _update(cls, self):
        if cls._is_train: cls.iou_mean = (1 - cls._momentum) * cls.iou_mean + \
                                         cls._momentum * self.iou.detach().mean().item()

    @classmethod
    def _scaled_loss(cls, self, gamma=1.9, delta=3):
        if isinstance(self.monotonous, bool):
            if self.monotonous:
                return (self.iou.detach() / self.iou_mean).sqrt()
            else:
                beta = self.iou.detach() / self.iou_mean
                alpha = delta * torch.pow(gamma, beta - delta)
                return beta / alpha
        return 1

def bbox_overlaps(pred: torch.Tensor,
                  target: torch.Tensor,
                  iou_mode: str = 'ciou',
                  bbox_format: str = 'xywh',
                  siou_theta: float = 4.0,
                  gamma: float = 0.5,
                  eps: float = 1e-7,
                  focal: bool = False,) -> torch.Tensor:
    r"""Calculate overlap between two set of bboxes.
    `Implementation of paper `Enhancing Geometric Factors into
    Model Learning and Inference for Object Detection and Instance
    Segmentation <https://arxiv.org/abs/2005.03572>`_.

    In the CIoU implementation of YOLOv5 and MMDetection, there is a slight
    difference in the way the alpha parameter is computed.

    mmdet version:
        alpha = (ious > 0.5).float() * v / (1 - ious + v)
    YOLOv5 version:
        alpha = v / (v - ious + (1 + eps)

    Args:
        pred (Tensor): Predicted bboxes of format (x1, y1, x2, y2)
            or (x, y, w, h),shape (n, 4).
        target (Tensor): Corresponding gt bboxes, shape (n, 4).
        iou_mode (str): Options are ('iou', 'ciou', 'giou', 'siou').
            Defaults to "ciou".
        bbox_format (str): Options are "xywh" and "xyxy".
            Defaults to "xywh".
        siou_theta (float): siou_theta for SIoU when calculate shape cost.
            Defaults to 4.0.
        eps (float): Eps to avoid log(0).

    Returns:
        Tensor: shape (n, ).
    """
    assert iou_mode in ('iou', 'ciou', 'giou', 'siou','diou','eiou','innersiou','innerciou','shapeiou','wiou')
    assert bbox_format in ('xyxy', 'xywh')
    if bbox_format == 'xywh':
        pred = HorizontalBoxes.cxcywh_to_xyxy(pred)
        target = HorizontalBoxes.cxcywh_to_xyxy(target)

    bbox1_x1, bbox1_y1 = pred[..., 0], pred[..., 1]
    bbox1_x2, bbox1_y2 = pred[..., 2], pred[..., 3]
    bbox2_x1, bbox2_y1 = target[..., 0], target[..., 1]
    bbox2_x2, bbox2_y2 = target[..., 2], target[..., 3]

    # Overlap
    overlap = (torch.min(bbox1_x2, bbox2_x2) -
               torch.max(bbox1_x1, bbox2_x1)).clamp(0) * \
              (torch.min(bbox1_y2, bbox2_y2) -
               torch.max(bbox1_y1, bbox2_y1)).clamp(0)

    # Union
    w1, h1 = bbox1_x2 - bbox1_x1, bbox1_y2 - bbox1_y1
    w2, h2 = bbox2_x2 - bbox2_x1, bbox2_y2 - bbox2_y1
    union = (w1 * h1) + (w2 * h2) - overlap + eps

    h1 = bbox1_y2 - bbox1_y1 + eps
    h2 = bbox2_y2 - bbox2_y1 + eps

    # IoU
    ious = overlap / union

    # enclose area
    enclose_x1y1 = torch.min(pred[..., :2], target[..., :2])
    enclose_x2y2 = torch.max(pred[..., 2:], target[..., 2:])
    enclose_wh = (enclose_x2y2 - enclose_x1y1).clamp(min=0)

    enclose_w = enclose_wh[..., 0]  # cw
    enclose_h = enclose_wh[..., 1]  # ch

    if iou_mode == 'ciou':
        # CIoU = IoU - ( (ρ^2(b_pred,b_gt) / c^2) + (alpha x v) )

        # calculate enclose area (c^2)
        enclose_area = enclose_w**2 + enclose_h**2 + eps

        # calculate ρ^2(b_pred,b_gt):
        # euclidean distance between b_pred(bbox2) and b_gt(bbox1)
        # center point, because bbox format is xyxy -> left-top xy and
        # right-bottom xy, so need to / 4 to get center point.
        rho2_left_item = ((bbox2_x1 + bbox2_x2) - (bbox1_x1 + bbox1_x2))**2 / 4
        rho2_right_item = ((bbox2_y1 + bbox2_y2) -
                           (bbox1_y1 + bbox1_y2))**2 / 4
        rho2 = rho2_left_item + rho2_right_item  # rho^2 (ρ^2)

        # Width and height ratio (v)
        wh_ratio = (4