求助：如何将自定义图片标注JSON与图片导入FiftyOne

问题描述 #

现有图片数据集和对应的标注JSON文件，尝试用FiftyOne的from_json或from_dir方法导入时报错，自行编写的代码运行耗时过长无法完成，代码如下：

import glob
import fiftyone as fo
import json

images_patt = "./data/*.jpg"

# Ex: load your custom label format
with open('./Waste_Plastics_result.json', 'r') as f:
    annotations = json.load(f)

# Create samples for your data
samples = []
for filepath in glob.glob(images_patt):
    sample = fo.Sample(filepath=filepath)

    # Convert detections to FiftyOne format
    detections = []
    for obj in annotations:
        width = obj["width"]
        height = obj["height"]
        #id = obj["id"]
        file_name = obj["file_name"]
        #image_id = obj["image_id"]
        category_id = obj["category_id"]
        metainfo_id = obj["metainfo_id"]
        # Bounding box coordinates should be relative values
        # in [0, 1] in the following format:
        # [top-left-x, top-left-y, width, height]
        bbox = obj["bbox"]
        #ignore = obj["ignore"]
        #iscrowd = obj["iscrowd"]
        area = obj["area"]

        detections.append(
            fo.Detection(width=width, height=height, file_name=file_name, category_id=category_id, metainfo_id=metainfo_id,
                         bbox=bbox, area=area)
        )

    # Store detections in a field name of your choice
    sample["ground_truth"] = fo.Detections(detections=detections)

    samples.append(sample)

# Create dataset
dataset = fo.Dataset.from_images_dir("./data")
dataset.add_samples(samples)

# FiftyOne session
session = fo.launch_app(dataset)
session.wait()  

代码问题分析 #

1. 嵌套循环逻辑错误：遍历每张图片时，把所有标注都添加到当前样本中，导致每个样本绑定了全量标注，数据冗余且计算量暴增，这是耗时过长的核心原因。
2. 重复创建数据集：先通过Dataset.from_images_dir导入所有图片，又调用add_samples重复添加样本，造成数据重复加载。
3. BBox格式不匹配：标注中的bbox是像素级绝对坐标，但FiftyOne的fo.Detection要求bbox为相对坐标（范围[0,1]），需要基于图片宽高转换。
4. 无效参数传入：fo.Detection不需要width、height、file_name这些图片属性，传入后会造成冗余数据。

修正后的代码 #

import fiftyone as fo
import json
import os

# 1. 加载标注并建立图片名到标注的映射
annotations_path = "./Waste_Plastics_result.json"
images_dir = "./data"

with open(annotations_path, 'r') as f:
    annotations = json.load(f)

# 按图片文件名分组标注
img_annot_map = {}
for ann in annotations:
    img_name = ann["file_name"]
    if img_name not in img_annot_map:
        img_annot_map[img_name] = []
    img_annot_map[img_name].append(ann)

# 2. 遍历图片，创建样本
samples = []
for img_name in os.listdir(images_dir):
    if not img_name.endswith(".jpg"):
        continue
    
    img_path = os.path.join(images_dir, img_name)
    sample = fo.Sample(filepath=img_path)
    
    # 获取当前图片的标注
    if img_name not in img_annot_map:
        # 无标注的图片也加入数据集，ground_truth为空
        sample["ground_truth"] = fo.Detections(detections=[])
        samples.append(sample)
        continue
    
    # 转换标注为FiftyOne格式
    detections = []
    for ann in img_annot_map[img_name]:
        # 获取图片宽高（从标注中读取，或从图片文件读取）
        img_width = ann["width"]
        img_height = ann["height"]
        
        # 转换绝对bbox为相对坐标：[x, y, w, h] -> [x/img_width, y/img_height, w/img_width, h/img_height]
        bbox_abs = ann["bbox"]
        bbox_rel = [
            bbox_abs[0] / img_width,
            bbox_abs[1] / img_height,
            bbox_abs[2] / img_width,
            bbox_abs[3] / img_height
        ]
        
        # 创建Detection对象，只保留必要字段
        det = fo.Detection(
            category_id=ann["category_id"],
            bbox=bbox_rel,
            area=ann["area"]
            # 如果有类别名称，可以添加label字段，比如label=category_names[ann["category_id"]]
        )
        detections.append(det)
    
    sample["ground_truth"] = fo.Detections(detections=detections)
    samples.append(sample)

# 3. 创建数据集并启动App
dataset = fo.Dataset("waste_plastics_dataset")
dataset.add_samples(samples)

session = fo.launch_app(dataset)
session.wait()

额外说明 #

• 如果标注中包含类别名称映射（比如categories字段），可以添加label字段到fo.Detection中，提升可视化效果。
• 若图片宽高在标注中不存在，可以用PIL库读取图片获取：

  from PIL import Image
  with Image.open(img_path) as img:
      img_width, img_height = img.size
  

内容的提问来源于stack exchange，提问作者user21840384