数据集质量控制

import hashlib
from collections import Counter
from pathlib import Path

from ultralytics.data.utils import IMG_FORMATS, check_det_dataset, img2label_paths

data = check_det_dataset("coco8.yaml")  # downloads COCO8 on first run; swap in your data YAML
names, seen = data["names"], {}

for split in ("train", "val"):
    images = sorted(p for p in Path(data[split]).rglob("*.*") if p.suffix[1:].lower() in IMG_FORMATS)
    instances, hashes, problems, background = Counter(), Counter(), [], 0
    for img, label in zip(images, img2label_paths([str(p) for p in images])):
        hashes[hashlib.md5(img.read_bytes()).hexdigest()] += 1
        rows = [r.split() for r in Path(label).read_text().splitlines() if r.strip()] if Path(label).exists() else []
        background += not rows
        if len(rows) != len({tuple(r) for r in rows}):
            problems.append(f"{img.name}: duplicate label rows")
        for r in rows:
            cls, w, h = int(r[0]), float(r[3]), float(r[4])
            if cls not in names:
                problems.append(f"{img.name}: class {cls} is not in names")
            elif w * h < 1e-4:
                problems.append(f"{img.name}: tiny {names[cls]} box")
            instances[cls] += 1
    seen[split] = set(instances)
    duplicates = sum(n - 1 for n in hashes.values())
    print(f"{split}: {len(images)} images, {background} background, {duplicates} duplicate images")
    for cls, n in instances.most_common():
        print(f"  {names.get(cls, cls):>12}  {n}")
    for problem in problems:
        print("  !", problem)

print("in val but never in train:", sorted(names[c] for c in seen["val"] - seen["train"] if c in names))

train: 4 images, 0 background, 0 duplicate images
        orange  4
          bowl  3
       giraffe  2
...
val: 4 images, 0 background, 0 duplicate images
        person  10
         horse  2
...
in val but never in train: ['dog', 'elephant', 'horse', 'person', 'umbrella']

from pathlib import Path

import torch

from ultralytics import YOLO
from ultralytics.data.utils import check_det_dataset, img2label_paths
from ultralytics.utils.metrics import box_iou
from ultralytics.utils.ops import xywh2xyxy

data = check_det_dataset("coco128.yaml")  # swap in your data YAML
model = YOLO("yolo26n.pt")  # a pretrained model, or your own first fine-tune
images = sorted(Path(data["train"]).glob("*.jpg"))

suspects = []
for result, label in zip(model(images, conf=0.6, stream=True, verbose=False), img2label_paths(map(str, images))):
    rows = [r.split() for r in Path(label).read_text().splitlines() if r.strip()] if Path(label).exists() else []
    boxes = xywh2xyxy(torch.tensor([[float(v) for v in r[1:5]] for r in rows]).reshape(-1, 4))
    iou = box_iou(result.boxes.xyxyn, boxes)  # model boxes vs labeled boxes
    for j, (cls, conf) in enumerate(zip(result.boxes.cls.int().tolist(), result.boxes.conf.tolist())):
        if not rows or iou[j].max() < 0.3:  # confident detection with no label anywhere near it
            suspects.append((conf, Path(result.path).name, model.names[cls]))

print(f"{len(suspects)} confident detections with no matching label; top 5:")
for conf, name, cls in sorted(suspects, reverse=True)[:5]:
    print(f"  {conf:.2f}  {cls:<12} {name}")

4 confident detections with no matching label; top 5:
  0.77  donut        000000000438.jpg
  0.75  person       000000000531.jpg
  0.71  donut        000000000438.jpg
  0.66  bottle       000000000436.jpg

   forklift  ████████████████████████ 8000 instances
   person    ████████████████ 5400
   pallet    ██ 700                              ← red flag

Dataset: forklift-v1   Audit date: 2026-05-09
Images: 4,820   Labels: 18,442   Background: 380 (7.9%)
Audit script: 3 invalid class IDs fixed, 41 duplicate label rows removed
Suspect review: 62 flagged, 19 missing labels added, 43 model errors ignored
Visual sample: 120 images, 4 mislabeled (3.3%), fixed
Split check: grouped by camera-day, every class in every split
Imbalance: pallet 2.5x below person; collecting more next round
Ready to train: yes