EricChenWei/neural-dqs-benchmark

Neural DQS Benchmark

96-variant COCO128 degradation benchmark for training and evaluating Neural Dataset Quality Score (Neural DQS) models.

This dataset contains feature vectors and ground-truth mAP@0.5 values for 96 systematically degraded versions of COCO128, used to validate the hypothesis:

DQS(D) ↑ ⟹ mAP(YOLO trained on D) ↑

Result: CV Pearson r = 0.929 (n=96, p<0.001)

---

Dataset Description

Each row represents one dataset variant. Features are extracted from the image/annotation set; map50 is the ground truth obtained by training YOLOv11n for 15 epochs.

Features

Column	Description	Range
annotation_quality (AQ)	0.6 × completeness + 0.4 × bbox geometry	[0, 1]
sharpness (IQ)	√(blur_score × noise_cleanliness)	[0, 1]
clip_diversity (CD)	Mean pairwise cosine distance in CLIP ViT-B/32 space	[0, 1]
lighting_diversity (LD)	Normalized brightness entropy (3 buckets)	[0, 1]
pose_diversity (PD)	Normalized aspect-ratio entropy	[0, 1]
class_balance (CB)	1 − Gini coefficient	[0, 1]
map50	mAP@0.5 — YOLOv11n trained 15 epochs (target variable)	[0, 1]
map50_95	mAP@0.5:0.95	[0, 1]

Degradation Categories (10 types)

Category	Variants	Description
Baseline	1	Original COCO128
Blur	20	Gaussian blur, kernel 3–61
Noise	8	Gaussian noise σ=2–100
Brightness	15	Factor 0.05–2.0
Label missing	13	10%–90% of label files blanked
Label noise	3	Bbox cx/cy shifted ±3%–20%
Combined	9	blur+dark, noise+dark, noise+blur (3 severities each)
Other	27	Dense sweeps across blur/brightness ranges

---

Key Result

Trained with Ridge(α=1.0) + PolynomialFeatures(degree=2):

Metric	Value
CV Pearson r (k=5)	0.929
CV R²	0.854
Train Pearson r	0.970

Top predictors:

• CD (CLIP Diversity): r = 0.892
• IQ (Image Quality): r = 0.661

---

Usage

import pandas as pd
from sklearn.linear_model import Ridge
from sklearn.preprocessing import StandardScaler, PolynomialFeatures
from sklearn.pipeline import Pipeline
from sklearn.model_selection import cross_val_predict, KFold
import numpy as np

df = pd.read_csv("hf://datasets/EricChenWei/neural-dqs-benchmark/dqs_training_data_v5.csv")

FEATURES = ["annotation_quality", "sharpness", "clip_diversity",
            "lighting_diversity", "pose_diversity", "class_balance"]

X = df[FEATURES].values
y = df["map50"].values

model = Pipeline([
    ("scaler", StandardScaler()),
    ("poly",   PolynomialFeatures(degree=2, include_bias=False)),
    ("ridge",  Ridge(alpha=1.0)),
])

cv = KFold(n_splits=5, shuffle=True, random_state=42)
y_cv = cross_val_predict(model, X, y, cv=cv)
r = np.corrcoef(y, y_cv)[0, 1]
print(f"CV Pearson r = {r:.4f}")  # → ~0.929

---

Related

• GitHub: ericchen931209/auto-dataset-builder
• Model: EricChenWei/neural-dqs

Citation

@software{chen2026adb,
  author  = {Chen, Yu-Wei},
  title   = {Auto Dataset Builder: An LLM-Assisted Framework for
             Automatic Dataset Construction with Neural Dataset Quality Scoring},
  year    = {2026},
  url     = {https://github.com/ericchen931209/auto-dataset-builder},
  license = {MIT}
}