Skip to content
教程阅读约需 5 分钟

使用 Potato 进行图像分类

在 Potato 中设置图像分类：用 instance_display 展示图像，用 radio 和 multiselect 方案收集标签，并加上质量检查。

Potato Team

大多数计算机视觉项目迟早都需要一堆标好的图像，而这些标注总得有人来做。无论是给产品照片分门别类、审核内容，还是给医学影像归类，Potato 负责界面，你只管标签。本教程涵盖几种常见配置：单标签和多标签分类、两级类目结构，以及让图像任务保持可靠的质量检查。

先把图像显示出来，再提问

在 Potato 里，分类和绘制是两件不同的事，把它们混为一谈是最常见的第一个错误。instance_display 负责把图片显示到屏幕上，annotation_schemes 负责提问。image_annotation 类型是用来在图像上画框和掩码的，分类任务完全用不到它。

yaml
annotation_task_name: "Product Image Classification"
task_dir: "."
 
data_files:
  - data/products.json
 
item_properties:
  id_key: product_id
  text_key: image_url
 
instance_display:
  fields:
    - key: image_url
      type: image
      label: "Product image"
      display_options:
        max_width: 600
        zoomable: true
 
annotation_schemes:
  - annotation_type: radio
    name: category
    description: "What type of product is shown?"
    labels:
      - Electronics
      - Clothing
      - Home & Garden
      - Sports & Outdoors
      - Books & Media
      - Other
    sequential_key_binding: true
 
output_annotation_dir: annotation_output/
export_annotation_format: jsonl

sequential_key_binding 把 1 到 6 这几个键分给六个标签。一个只需扫一眼、点一下的任务，速度基本就取决于键盘。

界面会在图像下方显示类别标签：

带有类别标签和键盘快捷键的图像分类界面图像标注界面，显示一张产品图片以及分类标签和键盘快捷键

数据格式

创建 data/products.json：

json
{"product_id": "P001", "image_url": "/images/products/laptop.jpg"}
{"product_id": "P002", "image_url": "/images/products/tshirt.jpg"}
{"product_id": "P003", "image_url": "/images/products/garden_tools.jpg"}
{"product_id": "P004", "image_url": "https://example.com/images/basketball.jpg"}

图像可以是任务目录下的本地路径，也可以是绝对 URL。text_key 指向存放路径的那个字段，Potato 会把这个字段交给显示组件。

多标签分类

一张图像同时属于好几个类别时，用 multiselect：

yaml
annotation_schemes:
  - annotation_type: multiselect
    name: attributes
    description: "Select all attributes that apply to this image"
    labels:
      - Contains people
      - Outdoor scene
      - Indoor scene
      - Contains text
      - Product photo
      - Lifestyle photo
      - Close-up shot
      - Wide angle

嵌套类目

Potato 没有给单选按钮提供嵌套标签类型，所以两级类目就是两个方案。每个方案都要有自己的 labels，而子类目列表必须覆盖所有父类目：

yaml
annotation_schemes:
  - annotation_type: radio
    name: main_category
    description: "Primary category"
    labels:
      - Apparel
      - Electronics
      - Home
 
  - annotation_type: radio
    name: subcategory
    description: "Subcategory within the primary category"
    labels:
      - Tops
      - Bottoms
      - Footwear
      - Phones & Tablets
      - Computers
      - Audio
      - Furniture
      - Kitchen
      - Decor

如果第二个列表会变得很长，hierarchical_multiselect 能把它收进一个控件，省掉拆成两个方案。

与分类并行的质量检查

yaml
annotation_schemes:
  - annotation_type: radio
    name: category
    description: "Product category"
    labels: [Electronics, Clothing, Home, Other]
 
  - annotation_type: radio
    name: image_quality
    description: "Is this image suitable for the product catalog?"
    labels:
      - name: Approved
        tooltip: "Clear, well-lit, product is visible"
      - name: Needs review
        tooltip: "Minor issues but potentially usable"
      - name: Rejected
        tooltip: "Poor quality, wrong product, or inappropriate"
 
  - annotation_type: multiselect
    name: quality_issues
    description: "If not approved, what issues are present?"
    labels:
      - Blurry or out of focus
      - Poor lighting
      - Wrong product shown
      - Watermark or text overlay
      - Inappropriate content
      - Multiple products in frame

针对单个标签的说明写在 tooltip 里，这是标签会读取的键。写成标签上的 description:，这段文字不会出现在标注者能看到的任何地方。

完整示例

yaml
annotation_task_name: "E-commerce Image Classification"
task_dir: "."
 
data_files:
  - data/images.json
 
item_properties:
  id_key: sku
  text_key: image_path
 
instance_display:
  fields:
    - key: image_path
      type: image
      label: "Product image"
      display_options:
        max_width: 700
        zoomable: true
 
annotation_schemes:
  - annotation_type: radio
    name: primary_category
    description: "Select the main product category"
    labels:
      - Electronics
      - Clothing & Accessories
      - Home & Living
      - Beauty & Personal Care
      - Sports & Fitness
      - Toys & Games
      - Other
    sequential_key_binding: true
 
  - annotation_type: multiselect
    name: image_attributes
    description: "Select all that apply"
    labels:
      - White background
      - Lifestyle shot
      - Multiple angles
      - Model wearing or using
      - Size reference included
      - Brand visible
    label_requirement:
      required: false
 
  - annotation_type: radio
    name: listing_ready
    description: "Is this image ready for the product listing?"
    labels:
      - Ready to publish
      - Needs editing
      - Reshoot required
    label_requirement:
      required: true
 
annotation_instructions: |
  ## Image Classification Guidelines
 
  ### Category selection
  Choose the most specific category that fits the product.
 
  ### Image quality
  - Ready to publish: clear, professional, meets standards
  - Needs editing: good photo that needs cropping or color correction
  - Reshoot required: fundamentally unsuitable
 
output_annotation_dir: annotation_output/
export_annotation_format: jsonl

图像标注的几点经验

不同显示器的颜色和亮度不一样，凡是这一点会影响判断的任务，都要确认标注者用的屏幕校准得还算靠谱。指南里放示例图片，别用文字去描述类别；一张图平息的争论比一整段话还多。把模糊图像该怎么处理写下来，因为模糊图像一定会出现。另外每隔一阵子让团队标同一批数据再做对比，标准漂移才能早点发现。

实现细节参见图像标注文档。

下一步

完整的图像标注文档参见图像标注。