目标检测trick实验和代码解析（基于yolov3）

实验结果

注：实验均不交叉，单独在base 上实验

ps	model	map	delta
baseline	001/19	84.75	0
focal loss a=0.25 r=2	002/41	86.16	+1.41
L2损失	003/19	85.11	+0.36
GIOU	004/17	84.85	+0.1
soft nms	001/19	84.09	-0.66

focal loss

参考：https://github.com/fizyr/keras-retinanet/blob/master/keras_retinanet/losses.py
发现$\alpha$分配给正负样本的不一样，一直以为$\alpha$是为了调节置信度损失和其他损失（比如回归损失）的比例的，果然看懂论文和看懂代码是两件不相干的事情
关键代码如下

alpha=0.25
gamma=2.0

#正样本的focal 系数
alpha_factor = tf.ones_like(respond_bbox) * alpha
alpha_factor = tf.where(tf.equal(respond_bbox, 1), alpha_factor, tf.zeros_like(respond_bbox))
focal_weight = tf.where(tf.equal(respond_bbox, 1), 1 - tf.nn.sigmoid(conv_raw_conf), tf.zeros_like(respond_bbox))
pos_focal_weight =  alpha_factor*focal_weight ** gamma

#负样本的focal系数
alpha_factor = tf.ones_like(respond_bgd) * alpha
alpha_factor = tf.where(tf.equal(respond_bgd, 1), 1-alpha_factor, tf.zeros_like(respond_bgd))
focal_weight = tf.where(tf.equal(respond_bgd, 1), tf.nn.sigmoid(conv_raw_conf), tf.zeros_like(respond_bgd))
neg_focal_weight =  alpha_factor*focal_weight ** gamma

conf_loss =  (
        pos_focal_weight*respond_bbox * tf.nn.sigmoid_cross_entropy_with_logits(labels=respond_bbox, logits=conv_raw_conf)
        +
        neg_focal_weight*respond_bgd * tf.nn.sigmoid_cross_entropy_with_logits(labels=respond_bbox, logits=conv_raw_conf)
)

L2 损失

类似Adaboost中的思想，如果预测值和真实值相差过大（难分样本），则应占有更大的比重，反之（易分样本）则是小比重

conf_loss = tf.pow(tf.abs(respond_bbox, pred_conf), 2)* (
                    respond_bbox * tf.nn.sigmoid_cross_entropy_with_logits(labels=respond_bbox, logits=conv_raw_conf)
                    +
                    respond_bgd * tf.nn.sigmoid_cross_entropy_with_logits(labels=respond_bbox, logits=conv_raw_conf)
            )

GIOU

def GIOU(boxes1, boxes2):
    boxes1_area = (boxes1[..., 2] - boxes1[..., 0]) * (boxes1[..., 3] - boxes1[..., 1])
    boxes2_area = (boxes2[..., 2] - boxes2[..., 0]) * (boxes2[..., 3] - boxes2[..., 1])

    # 计算出boxes1与boxes1相交部分的左上角坐标、右下角坐标
    intersection_left_up = tf.maximum(boxes1[..., :2], boxes2[..., :2])
    intersection_right_down = tf.minimum(boxes1[..., 2:], boxes2[..., 2:])

    # 因为两个boxes没有交集时，(right_down - left_up) < 0，所以maximum可以保证当两个boxes没有交集时，它们之间的iou为0
    intersection = tf.maximum(intersection_right_down - intersection_left_up, 0.0)
    inter_area = intersection[..., 0] * intersection[..., 1]
    union_area = boxes1_area + boxes2_area - inter_area
    IOU = 1.0 * inter_area / union_area
	#最小闭包 左上角坐标和右下角坐标
    enclose_left_up = tf.minimum(boxes1[..., :2], boxes2[..., :2])
    enclose_right_down = tf.maximum(boxes1[..., 2:], boxes2[..., 2:])
    enclose = tf.maximum(enclose_right_down - enclose_left_up, 0.0)
    enclose_area = enclose[..., 0] * enclose[..., 1]
    GIOU = IOU - 1.0 * (enclose_area - union_area) / enclose_area
    return GIOU

NMS / SOFT NMS

def nms(bboxes, score_threshold, iou_threshold, sigma=0.3, method='nms'):
    """
    :param bboxes:
    假设有N个bbox的score大于score_threshold，那么bboxes的shape为(N, 6)，存储格式为(xmin, ymin, xmax, ymax, score, class)
    其中(xmin, ymin, xmax, ymax)的大小都是相对于输入原图的，score = conf * prob，class是bbox所属类别的索引号
    :return: best_bboxes
    假设NMS后剩下N个bbox，那么best_bboxes的shape为(N, 6)，存储格式为(xmin, ymin, xmax, ymax, score, class)
    其中(xmin, ymin, xmax, ymax)的大小都是相对于输入原图的，score = conf * prob，class是bbox所属类别的索引号
    """
    classes_in_img = list(set(bboxes[:, 5]))
    best_bboxes = []
	#遍历每个类
    for cls in classes_in_img:
        cls_mask = (bboxes[:, 5] == cls)
        cls_bboxes = bboxes[cls_mask]
        while len(cls_bboxes) > 0:
            max_ind = np.argmax(cls_bboxes[:, 4])
            best_bbox = cls_bboxes[max_ind] #拿到score最高的box
            best_bboxes.append(best_bbox)
            cls_bboxes = np.concatenate([cls_bboxes[: max_ind], cls_bboxes[max_ind + 1:]]) #去除最高score的其他box
            iou = iou_calc1(best_bbox[np.newaxis, :4], cls_bboxes[:, :4])
            assert method in ['nms', 'soft-nms']
            weight = np.ones((len(iou),), dtype=np.float32)
            if method == 'nms': #使用nms
                iou_mask = iou > iou_threshold
                weight[iou_mask] = 0.0
            if method == 'soft-nms': #使用softnms
                weight = np.exp(-(1.0 * iou ** 2 / sigma))
            cls_bboxes[:, 4] = cls_bboxes[:, 4] * weight
            score_mask = cls_bboxes[:, 4] > score_threshold
            cls_bboxes = cls_bboxes[score_mask]
    return best_bboxes