如何在某些图片上上传重复标签以实现 azure 自定义视觉？答案

【问题标题】：How to uploading duplicate tags at some picture for azure custom vision?如何在某些图片上上传重复标签以实现 azure 自定义视觉？
【发布时间】：2019-08-15 06:38:15
【问题描述】：

我有一个关于 azure custom vision 的问题。我有一个用于对象检测的自定义视觉项目。我使用 python SDK 创建项目（参见：https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/python-tutorial-od）。但是我在上传的过程中发现了一些错误。例如，有一张图片在这张图片中有 3 个人。所以我在这张照片中标记了 3 个相同班级的“人”。但上传后，我只是在自定义视觉网站上发现这张图片中标记了 1 个“人”。但是其他类也可以，比如这张图片也可以有“人”、“汽车”和“踏板车”。看起来图片上只能有一个相同的班级。

我尝试使用python SDK（见：https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/python-tutorial-od）上传我的图片和标签信息。

A0_tag = trainer.create_tag(project.id, "A0")
A1_tag = trainer.create_tag(project.id, "A1")
A2_tag = trainer.create_tag(project.id, "A2")

A0_image_regions={
"0001.jpg":[0.432291667,0.28125,0.080729167,0.09765625],
"0001.jpg":[0.34765625,0.385742188,0.131510417,0.135742188],
"0001.jpg":[0.479166667,0.385742188,0.130208333,0.135742188],
"0003.jpg":[0.19921875,0.158203125,0.083333333,0.099609375]
}

上面的代码可以看到我在0001.jpg中上传了三个“A0”类。但是在网站的GUI界面中，我最终只能看到0001.jpg上面存在一个“A0”类。有什么办法可以解决这个问题吗？

基于 cthrash 代码。我对代码进行了一些更改以使其正常工作。这是修改后的代码：

A0_tag = trainer.create_tag(project.id, "TestA")
A1_tag = trainer.create_tag(project.id, "TestB")
A2_tag = trainer.create_tag(project.id, "TestC")

A0_image_regions = {
    A0_tag.id : [
                ("2300.png",[0.787109375,0.079681275,0.068359375,0.876494024]),
                ("0920.png",[0.2109375,0.065737052,0.059570313,0.892430279]),
                ("0920.png",[0.291015625,0.061752988,0.05859375,0.894422311]),
    ]
}

A1_image_regions = {
        A1_tag.id : [
                    ("2000.png",[0.067382813,0.073705179,0.030273438,0.878486056]),
                    ("2000.png",[0.126953125,0.075697211,0.030273438,0.878486056]),
                    ("2000.png",[0.184570313,0.079681275,0.030273438,0.878486056]),
                    ("2000.png",[0.232421875,0.079681275,0.030273438,0.878486056]),
    ],
}

A2_image_regions = {
        A2_tag.id : [
                ("1400.png",[0.649414063,0.065737052,0.104492188,0.894422311]),
                ("2300.png",[0.602539063,0.061752988,0.106445313,0.892430279]),
                ("0920.png",[0.634765625,0.067729084,0.124023438,0.88247012]),
                ("0800.png",[0.579101563,0.06374502,0.04296875,0.888446215]),
    ],
}



regions_map = {}
for tag_id in A0_image_regions:
    for filename,[x,y,w,h] in A0_image_regions[tag_id]:
        regions = regions_map.get(filename,[])
        regions.append(Region(tag_id=A0_tag.id, left=x, top=y, width=w, height=h))
        regions_map[filename] = regions

for tag_id in A1_image_regions:
     for filename,[x,y,w,h] in A1_image_regions[tag_id]:
        regions = regions_map.get(filename,[])
        regions.append(Region(tag_id=A1_tag.id, left=x, top=y, width=w, height=h))
        regions_map[filename] = regions


for tag_id in A2_image_regions:
     for filename,[x,y,w,h] in A2_image_regions[tag_id]:
        regions = regions_map.get(filename,[])
        regions.append(Region(tag_id=A2_tag.id, left=x, top=y, width=w, height=h))
        regions_map[filename] = regions




tagged_images_with_regions = []
for filename in regions_map:
    regions = regions_map[filename]
    with open("<your path>" + filename, mode="rb") as image_contents:



        tagged_images_with_regions.append(ImageFileCreateEntry(name=filename, contents=image_contents.read(), regions=regions))
upload_result = trainer.create_images_from_files(project.id, images=tagged_images_with_regions)

【问题讨论】：

标签： python azure microsoft-custom-vision

【解决方案1】：

您已经创建了A0_image_regions，但只要您对任何给定图像有多个边界框，就会覆盖该键。所以这是行不通的。

但也许更重要的是，您需要以图像为主要对象来调用训练器，并将所有相关的图像区域集中在一起。换句话说，在您的示例中，0001.jpg 具有三个A0 实例，但它也可能具有A1 和/或A2 的实例，并且这需要是单个ImageFile 条目。因此，我将按照以下内容修改示例：

A0_tag = trainer.create_tag(project.id, "A0")
A1_tag = trainer.create_tag(project.id, "A1")
A2_tag = trainer.create_tag(project.id, "A2")

image_regions = {
    A0_tag.id : [
        ("0001.jpg", [0.432291667,0.28125,0.080729167,0.09765625]),
        ("0001.jpg", [0.34765625,0.385742188,0.131510417,0.135742188]),
        ("0001.jpg", [0.479166667,0.385742188,0.130208333,0.135742188]),
        ("0003.jpg", [0.19921875,0.158203125,0.083333333,0.099609375])
    ],
    A1_tag.id : [] # add images/bounding boxes for A1
    A2_tag.id : [] # add images/bounding boxes for A2
}

regions_map = {}
for tag_id in image_regions:
    for filename,[x,y,w,h] in image_regions[tag_id]:
        regions = regions_map.get(filename,[])
        regions.append(Region(tag_id, left=x, top=y, width=w, height=h))
        regions_map[filename] = regions

tagged_images_with_regions = []
for filename in regions_map:
    regions = regions_map[filename]
    with open(base_image_url + filename, mode="rb") as image_contents:
        tagged_images_with_regions.append(ImageFileCreateEntry(name=filename, contents=image_contents.read(), regions=regions))

upload_result = trainer.create_images_from_files(project.id, images=tagged_images_with_regions)

【讨论】：

感谢您的回复，这些代码可以使用。我做了一些修改以使循环工作。代码保存在我的问题后面。

【解决方案2】：

听起来您只想为一张图片中的 3 个人标记一个标签 person，但这没有意义，不是问题。实际上，标签是针对图片进行标记的，而不是针对图片中显示人的像素区域。

因此，标签person 仅有助于检测在训练模型后至少有一个人的事实，而不是car 或scooter。如果要检测不同的人，则需要为图片中的三个不同的人添加person1、person2和person3三个标签。

请参阅 wiki 页面Object detection 及其参考链接，了解有关机器学习和深度学习原理的更多详细信息。

【讨论】：

【解决方案3】：

如果您没有更改示例代码中的其他任何内容，那么它会尝试将带有一个边界框的图像“0.001.jpg”上传三次，最后两次上传失败，因为它们与您的第一个重复图像上传图片。

“0.001.jpg”请上传一次，带三个边框，或者先上传图片再上传三个边框。

【讨论】：