目标检测yolo, voc, coco的BBox格式转换

yolo, voc, coco bbox格式互转函数

yolo: [xmid, ymid, w, h],归一化到0-1

voc: [x1, y1, x2, y2]

coco: [xmin, ymin, w, h]

 

def voc2yolo(bboxes, image_height=720, image_width=1280):
    """
    voc  => [x1, y1, x2, y2]
    yolo => [xmid, ymid, w, h] (normalized)
    """
    
    bboxes = bboxes.copy().astype(float) # otherwise all value will be 0 as voc_pascal dtype is np.int
    
    bboxes[..., [0, 2]] = bboxes[..., [0, 2]]/ image_width
    bboxes[..., [1, 3]] = bboxes[..., [1, 3]]/ image_height
    
    w = bboxes[..., 2] - bboxes[..., 0]
    h = bboxes[..., 3] - bboxes[..., 1]
    
    bboxes[..., 0] = bboxes[..., 0] + w/2
    bboxes[..., 1] = bboxes[..., 1] + h/2
    bboxes[..., 2] = w
    bboxes[..., 3] = h
    
    return bboxes

def yolo2voc(bboxes, image_height=720, image_width=1280):
    """
    yolo => [xmid, ymid, w, h] (normalized)
    voc  => [x1, y1, x2, y2]
    
    """ 
    bboxes = bboxes.copy().astype(float) # otherwise all value will be 0 as voc_pascal dtype is np.int
    
    bboxes[..., [0, 2]] = bboxes[..., [0, 2]]* image_width
    bboxes[..., [1, 3]] = bboxes[..., [1, 3]]* image_height
    
    bboxes[..., [0, 1]] = bboxes[..., [0, 1]] - bboxes[..., [2, 3]]/2
    bboxes[..., [2, 3]] = bboxes[..., [0, 1]] + bboxes[..., [2, 3]]
    
    return bboxes

def coco2yolo(bboxes, image_height=720, image_width=1280):
    """
    coco => [xmin, ymin, w, h]
    yolo => [xmid, ymid, w, h] (normalized)
    """
    
    bboxes = bboxes.copy().astype(float) # otherwise all value will be 0 as voc_pascal dtype is np.int
    
    # normolizinig
    bboxes[..., [0, 2]]= bboxes[..., [0, 2]]/ image_width
    bboxes[..., [1, 3]]= bboxes[..., [1, 3]]/ image_height
    
    # converstion (xmin, ymin) => (xmid, ymid)
    bboxes[..., [0, 1]] = bboxes[..., [0, 1]] + bboxes[..., [2, 3]]/2
    
    return bboxes

def yolo2coco(bboxes, image_height=720, image_width=1280):
    """
    yolo => [xmid, ymid, w, h] (normalized)
    coco => [xmin, ymin, w, h]
    
    """ 
    bboxes = bboxes.copy().astype(float) # otherwise all value will be 0 as voc_pascal dtype is np.int
    
    # denormalizing
    bboxes[..., [0, 2]]= bboxes[..., [0, 2]]* image_width
    bboxes[..., [1, 3]]= bboxes[..., [1, 3]]* image_height
    
    # converstion (xmid, ymid) => (xmin, ymin) 
    bboxes[..., [0, 1]] = bboxes[..., [0, 1]] - bboxes[..., [2, 3]]/2
    
    return bboxes

def voc2coco(bboxes, image_height=720, image_width=1280):
    bboxes  = voc2yolo(bboxes, image_height, image_width)
    bboxes  = yolo2coco(bboxes, image_height, image_width)
    return bboxes

版权声明:本文为CSDN博主「fwu11」的原创文章,遵循CC 4.0 BY-SA版权协议,转载请附上原文出处链接及本声明。
原文链接:https://blog.csdn.net/weixin_42552135/article/details/122533582

fwu11

我还没有学会写个人说明!

暂无评论

发表评论

相关推荐

目标检测中的先验框(Anchor)

什么是先验框?
了解过目标检测算法的朋友们肯定知道先验框(Anchor)的概念,那么什么是先验框,为什么要有先验框?若要解释这个问题,首先我们需要了解边界框回归原理。
b

YOLOv3学习——锚框和候选区域

YOLOv3学习之锚框和候选区域
单阶段目标检测模型YOLOv3
R-CNN系列算法需要先产生候选区域,再对候选区域做分类和位置坐标的预测,这类算法被称为两阶段目标检测算法。近几年,很多研究人员相继