#!/usr/bin/env python3 """ Script de vérification de la qualité du dataset """ import json import cv2 import numpy as np from pathlib import Path import argparse from collections import defaultdict import matplotlib.pyplot as plt class DatasetValidator: def __init__(self, annotations_file, images_dir): self.annotations_file = Path(annotations_file) self.images_dir = Path(images_dir) self.errors = [] self.warnings = [] self.stats = defaultdict(int) def load_annotations(self): """Charge les annotations COCO""" with open(self.annotations_file, 'r') as f: return json.load(f) def validate(self): """Valide le dataset""" print("🔍 Validation du dataset...") print("=" * 60) data = self.load_annotations() # Statistiques de base self.stats['num_images'] = len(data['images']) self.stats['num_annotations'] = len(data['annotations']) self.stats['num_classes'] = len(data['categories']) print(f"📊 Statistiques de base:") print(f" Images: {self.stats['num_images']}") print(f" Annotations: {self.stats['num_annotations']}") print(f" Classes: {self.stats['num_classes']}") # Valider les images self.validate_images(data) # Valider les annotations self.validate_annotations(data) # Analyser la distribution self.analyze_distribution(data) # Rapport final self.print_report() def validate_images(self, data): """Valide les images""" print("\n📸 Validation des images...") for img_info in data['images']: img_path = self.images_dir / img_info['file_name'] # Vérifier existence if not img_path.exists(): self.errors.append(f"Image manquante: {img_info['file_name']}") continue # Vérifier dimensions img = cv2.imread(str(img_path)) if img is None: self.errors.append(f"Image corrompue: {img_info['file_name']}") continue h, w = img.shape[:2] if h != img_info['height'] or w != img_info['width']: self.warnings.append( f"Dimensions incorrectes: {img_info['file_name']} " f"(déclaré: {img_info['width']}x{img_info['height']}, " f"réel: {w}x{h})" ) # Vérifier résolution minimale if w < 640 or h < 480: self.warnings.append( f"Résolution faible: {img_info['file_name']} ({w}x{h})" ) self.stats['total_pixels'] += w * h def validate_annotations(self, data): """Valide les annotations""" print("🏷️ Validation des annotations...") # Grouper par image annotations_by_image = defaultdict(list) for anno in data['annotations']: annotations_by_image[anno['image_id']].append(anno) # Statistiques par classe class_counts = defaultdict(int) bbox_sizes = [] for img_info in data['images']: img_id = img_info['id'] annos = annotations_by_image.get(img_id, []) # Images sans annotations if len(annos) == 0: self.stats['images_without_annotations'] += 1 for anno in annos: class_counts[anno['category_id']] += 1 # Valider bbox x, y, w, h = anno['bbox'] # Bbox hors limites if x < 0 or y < 0: self.errors.append( f"Bbox hors limites (négatif): image {img_id}" ) if x + w > img_info['width'] or y + h > img_info['height']: self.errors.append( f"Bbox hors limites (dépassement): image {img_id}" ) # Bbox trop petite if w < 10 or h < 10: self.warnings.append( f"Bbox très petite ({w}x{h}): image {img_id}" ) # Bbox trop grande (probablement erreur) img_area = img_info['width'] * img_info['height'] bbox_area = w * h if bbox_area / img_area > 0.5: self.warnings.append( f"Bbox très grande (>50% image): image {img_id}" ) bbox_sizes.append((w, h)) # Statistiques self.stats['class_counts'] = dict(class_counts) self.stats['bbox_sizes'] = bbox_sizes # Moyenne d'annotations par image if self.stats['num_images'] > 0: self.stats['avg_annotations_per_image'] = \ self.stats['num_annotations'] / self.stats['num_images'] def analyze_distribution(self, data): """Analyse la distribution du dataset""" print("\n📈 Analyse de la distribution...") # Distribution par classe print("\n Distribution par classe:") for class_id, count in self.stats['class_counts'].items(): class_name = next( (c['name'] for c in data['categories'] if c['id'] == class_id), f"Unknown ({class_id})" ) percentage = (count / self.stats['num_annotations']) * 100 print(f" {class_name}: {count} ({percentage:.1f}%)") # Taille moyenne des bbox if self.stats['bbox_sizes']: widths = [w for w, h in self.stats['bbox_sizes']] heights = [h for w, h in self.stats['bbox_sizes']] print(f"\n Taille des bbox:") print(f" Largeur moyenne: {np.mean(widths):.1f}px") print(f" Hauteur moyenne: {np.mean(heights):.1f}px") print(f" Largeur médiane: {np.median(widths):.1f}px") print(f" Hauteur médiane: {np.median(heights):.1f}px") def print_report(self): """Affiche le rapport final""" print("\n" + "=" * 60) print("📋 RAPPORT FINAL") print("=" * 60) # Résumé print(f"\n✅ Images validées: {self.stats['num_images']}") print(f"✅ Annotations validées: {self.stats['num_annotations']}") print(f"📊 Moyenne annotations/image: {self.stats.get('avg_annotations_per_image', 0):.2f}") if self.stats.get('images_without_annotations', 0) > 0: pct = (self.stats['images_without_annotations'] / self.stats['num_images']) * 100 print(f"⚠️ Images sans annotation: {self.stats['images_without_annotations']} ({pct:.1f}%)") # Erreurs if self.errors: print(f"\n❌ ERREURS ({len(self.errors)}):") for error in self.errors[:10]: # Limiter à 10 print(f" • {error}") if len(self.errors) > 10: print(f" ... et {len(self.errors) - 10} autres") else: print("\n✅ Aucune erreur détectée") # Avertissements if self.warnings: print(f"\n⚠️ AVERTISSEMENTS ({len(self.warnings)}):") for warning in self.warnings[:10]: print(f" • {warning}") if len(self.warnings) > 10: print(f" ... et {len(self.warnings) - 10} autres") else: print("\n✅ Aucun avertissement") # Recommandations print("\n💡 RECOMMANDATIONS:") if self.stats['num_images'] < 500: print(" ⚠️ Dataset petit (<500 images), collectez plus de données") elif self.stats['num_images'] < 1000: print(" ℹ️ Dataset correct, 1000+ images recommandé pour production") else: print(" ✅ Taille du dataset excellente") if self.stats.get('images_without_annotations', 0) / self.stats['num_images'] > 0.1: print(" ⚠️ Trop d'images sans annotation (>10%), vérifier la qualité") # Score global score = 100 score -= len(self.errors) * 5 score -= len(self.warnings) * 1 score = max(0, score) print(f"\n🎯 Score de qualité: {score}/100") if score >= 90: print(" ✅ Excellent dataset, prêt pour l'entraînement") elif score >= 70: print(" ⚠️ Dataset acceptable, corrections recommandées") else: print(" ❌ Dataset problématique, corrections nécessaires") def visualize_distribution(self): """Crée des visualisations de la distribution""" if not self.stats.get('bbox_sizes'): return fig, axes = plt.subplots(2, 2, figsize=(12, 10)) # Distribution des largeurs widths = [w for w, h in self.stats['bbox_sizes']] axes[0, 0].hist(widths, bins=30, edgecolor='black') axes[0, 0].set_title('Distribution des largeurs de bbox') axes[0, 0].set_xlabel('Largeur (px)') axes[0, 0].set_ylabel('Fréquence') # Distribution des hauteurs heights = [h for w, h in self.stats['bbox_sizes']] axes[0, 1].hist(heights, bins=30, edgecolor='black') axes[0, 1].set_title('Distribution des hauteurs de bbox') axes[0, 1].set_xlabel('Hauteur (px)') axes[0, 1].set_ylabel('Fréquence') # Scatter largeur vs hauteur axes[1, 0].scatter(widths, heights, alpha=0.5) axes[1, 0].set_title('Largeur vs Hauteur') axes[1, 0].set_xlabel('Largeur (px)') axes[1, 0].set_ylabel('Hauteur (px)') # Distribution par classe if self.stats.get('class_counts'): classes = list(self.stats['class_counts'].keys()) counts = list(self.stats['class_counts'].values()) axes[1, 1].bar(classes, counts) axes[1, 1].set_title('Distribution par classe') axes[1, 1].set_xlabel('Classe') axes[1, 1].set_ylabel('Nombre d\'annotations') plt.tight_layout() plt.savefig('dataset_distribution.png', dpi=150) print("\n📊 Graphiques sauvegardés: dataset_distribution.png") def main(): parser = argparse.ArgumentParser( description='Validation de la qualité du dataset' ) parser.add_argument('--annotations', type=str, required=True, help='Fichier annotations.json') parser.add_argument('--images', type=str, required=True, help='Dossier des images') parser.add_argument('--visualize', action='store_true', help='Créer des visualisations') args = parser.parse_args() validator = DatasetValidator(args.annotations, args.images) validator.validate() if args.visualize: validator.visualize_distribution() if __name__ == '__main__': main()