Azure语义分割RLE转Mask异常：解码后仅显示右上角竖线求助

我用Azure Data Labelling制作了Semantic Segmentation (Preview)数据集，导出到本地后有images文件夹的图片和JSON标签文件。尝试解码RLE字符串生成分割Mask时，结果始终只有右上角一条竖线，使用的解码代码如下：

JSON中的RLE数据 #

{"rle":[121256,0,3,1,715,0,6,1,712,0,8,1,712,0,8,1,712,0,9,1,711,0,12,1,708,0,19,1,702,0,21,1,699,0,23,1,698,0,22,1,699,0,21,1,700,0,20,1,704,0,15,1,675,0,5,1,714,0,6,1,705,0,4,1,4,0,8,1,703,0,18,1,701,0,20,1,699,0,22,1,697,0,24,1,696,0,25,1,694,0,26,1,693,0,27,1,693,0,28,1,692,0,28,1,692,0,29,1,691,0,29,1,690,0,30,1,691,0,30,1,690,0,30,1,690,0,30,1,690,0,30,1,690,0,30,1,690,0,30,1,690,0,31,1,689,0,31,1,689,0,31,1,689,0,30,1,691,0,29,1,691,0,29,1,691,0,28,1,692,0,28,1,693,0,27,1,693,0,27,1,694,0,26,1,695,0,25,1,696,0,23,1,698,0,22,1,700,0,20,1,701,0,19,1,702,0,18,1,704,0,16,1,705,0,15,1,707,0,13,1,708,0,12,1,710,0,10,1,711,0,9,1,713,0,7,1,714,0,6,1,355398,0]}

使用的解码代码 #

import json
import numpy as np
import cv2

# rle_decode function
def rle_decode(rle_list, shape, fill_value=1, dtype=int, relative=False):
    rle_str = ' '.join(map(str, rle_list))
    s = rle_str.strip().split(" ")
    starts, lengths = np.array([np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])])
    mask = np.zeros(np.prod(shape), dtype=dtype)
    if relative:
        start = 0
        for index, length in zip(starts, lengths):
            start = start + index
            end = start + length
            mask[start:end] = fill_value
            start = end
        return mask.reshape(shape[::-1]).T
    else:
        starts -= 1
        ends = starts + lengths
        for lo, hi in zip(starts, ends):
            mask[lo:hi] = fill_value
        return mask.reshape(shape[::-1]).T

# Read JSON file
with open('downloads/labels/2D82C4F8-F1F2-4042-AE45-C4988A55048A.json', 'r') as f:
    data = json.load(f)

rle_list = data['rle']
shape = (100, 100)  # Assuming the shape is known or provided

# Decode RLE list
mask = rle_decode(rle_list, shape)

# Create a black image
image = np.zeros(shape, dtype=np.uint8)

# Draw the mask on the black image
image[mask == 1] = 255  # White color for mask

# Display the image
cv2.imshow('Mask', image)
cv2.waitKey(0)
cv2.destroyAllWindows()

当前输出仅为右上角一条竖线，需要正确解码RLE得到完整分割Mask用于模型训练。

问题出在两个核心点：

1. 错误的图像尺寸假设：手动设置的(100,100)和实际图像尺寸不匹配。从RLE的第一个数值121256来看，这个数值远大于100*100=10000，说明实际图像的像素总数远大于10000。

   • 解决：从对应的图片文件中读取真实尺寸，比如用cv2.imread获取图像的高和宽：

     img = cv2.imread('path/to/your/image.jpg')
     shape = (img.shape[0], img.shape[1])  # (height, width)
     

2. Azure Data Labelling的RLE编码规则：Azure导出的语义分割RLE是相对编码（即每个起始位置是相对于前一段结束的偏移量），但原代码默认使用了relative=False的绝对编码逻辑，导致解码位置完全错误。

   • 解决：调用解码函数时必须匹配相对编码逻辑，同时Azure的RLE按**行优先（row-major）**排列，不需要额外转置。

修正后的完整代码 #

import json
import numpy as np
import cv2

def rle_decode(rle_list, shape, fill_value=1, dtype=int):
    # 适配Azure的相对偏移RLE编码，行优先排列
    mask = np.zeros(np.prod(shape), dtype=dtype)
    current_pos = 0
    # RLE列表为[偏移量, 长度]交替结构，处理末尾的填充偏移
    for i in range(0, len(rle_list), 2):
        offset = rle_list[i]
        length = rle_list[i+1]
        if length == 0:
            current_pos += offset
            continue
        start = current_pos + offset
        end = start + length
        mask[start:end] = fill_value
        current_pos = end
    # 按(height, width)直接reshape，无需转置
    return mask.reshape(shape)

# 读取JSON标签文件
with open('downloads/labels/2D82C4F8-F1F2-4042-AE45-C4988A55048A.json', 'r') as f:
    data = json.load(f)

rle_list = data['rle']
# 读取对应图片获取真实尺寸（替换为你的图片路径）
img_path = 'downloads/images/2D82C4F8-F1F2-4042-AE45-C4988A55048A.jpg'
img = cv2.imread(img_path)
shape = (img.shape[0], img.shape[1])  # (height, width)

# 解码RLE生成Mask
mask = rle_decode(rle_list, shape)

# 可视化正确的Mask
mask_img = (mask * 255).astype(np.uint8)
cv2.imshow('Correct Mask', mask_img)
cv2.waitKey(0)
cv2.destroyAllWindows()

额外验证点 #

• 确认JSON中的rle列表是成对的（偏移量+长度），Azure导出的RLE最后通常会以[大偏移量, 0]结尾，用于填充到图像末尾，修正后的代码已处理这种情况。
• 确保图片路径和标签文件名对应，Azure导出的标签文件名通常和图片文件名一致（仅后缀不同）。

内容的提问来源于stack exchange，提问作者Muhammad Faizan