#!/usr/bin/env python3 """ OV-COCO Failure Case Analysis - Cross-Model Comparison V2 开放词汇检测（OVD）场景下的模型对比分析。 核心逻辑（针对OVD任务优化）： 1. 对于每个GT目标（novel类别 + 小目标），在所有预测中找IoU最高的框（不限类别） 2. 判断该预测框的分类是否正确（预测类别 == GT类别） 3. DeCLIP优势case = DeCLIP分类正确 + CLIPSelf/CLIP分类错误 这样可以清晰展示DeCLIP在novel小目标上的分类优势。 Usage: python compare_models_v2.py \ --declip-pred results/declip_vitb16/predictions.json \ --clipself-pred results/clipself_vitb16/predictions.json \ --clip-pred results/clearclip_vitb16/predictions.json \ --ann-file /path/to/instances_val2017_all_2.json \ --seen-classes /path/to/mscoco_seen_classes.json \ --output analysis_output/comparison_vitb16 """ import argparse import json from collections import defaultdict from pathlib import Path from typing import Dict, List import numpy as np from pycocotools.coco import COCO # COCO目标大小标准 SIZE_THRESHOLDS = { 'small': (0, 32 * 32), 'medium': (32 * 32, 96 * 96), 'large': (96 * 96, float('inf')) } def parse_args(): parser = argparse.ArgumentParser(description='Cross-model comparison analysis V2') parser.add_argument('--declip-pred', required=True, help='DeCLIP predictions JSON') parser.add_argument('--clipself-pred', required=True, help='CLIPSelf predictions JSON') parser.add_argument('--clip-pred', required=True, help='CLIP (ClearCLIP) predictions JSON') parser.add_argument('--ann-file', required=True, help='COCO annotation file') parser.add_argument('--seen-classes', required=True, help='Seen classes JSON file') parser.add_argument('--output', required=True, help='Output directory') parser.add_argument('--iou-threshold', type=float, default=0.5, help='IoU threshold for TP') parser.add_argument('--top-k', type=int, default=50, help='Number of top cases to output') return parser.parse_args() def compute_iou(bbox1: List[float], bbox2: List[float]) -> float: """计算IoU，bbox格式: [x, y, w, h]""" x1, y1, w1, h1 = bbox1 x2, y2, w2, h2 = bbox2 box1 = [x1, y1, x1 + w1, y1 + h1] box2 = [x2, y2, x2 + w2, y2 + h2] xi1 = max(box1[0], box2[0]) yi1 = max(box1[1], box2[1]) xi2 = min(box1[2], box2[2]) yi2 = min(box1[3], box2[3]) inter_area = max(0, xi2 - xi1) * max(0, yi2 - yi1) box1_area = w1 * h1 box2_area = w2 * h2 union_area = box1_area + box2_area - inter_area return inter_area / union_area if union_area > 0 else 0 def get_size_category(area: float) -> str: """根据面积返回大小类别""" for size_name, (min_area, max_area) in SIZE_THRESHOLDS.items(): if min_area <= area < max_area: return size_name return 'large' def load_predictions(pred_file: str) -> Dict[int, List[dict]]: """ 加载预测结果，按image_id组织（包含所有类别的预测） Returns: Dict[image_id -> List[prediction]] """ with open(pred_file, 'r') as f: predictions = json.load(f) # 按image_id组织，并按score降序排列 pred_by_img = defaultdict(list) for pred in predictions: pred_by_img[pred['image_id']].append(pred) # 按score降序排列 for img_id in pred_by_img: pred_by_img[img_id] = sorted( pred_by_img[img_id], key=lambda x: x['score'], reverse=True ) return dict(pred_by_img) def find_best_matching_pred( predictions: List[dict], gt_bbox: List[float], gt_category_id: int, id_to_name: dict, min_iou: float = 0.1 ) -> dict: """ 在所有预测中找与GT位置最匹配的框（IoU最高），不限类别。 然后判断分类是否正确。 Args: predictions: 该图像的所有预测 gt_bbox: GT的bbox gt_category_id: GT的类别ID id_to_name: category_id到name的映射 min_iou: 最小IoU阈值（低于此值认为没有定位到） Returns: dict: { 'pred': 最佳匹配的预测框（或None）, 'iou': IoU值, 'localized': 是否定位成功（IoU >= min_iou）, 'classified_correct': 分类是否正确, 'pred_category_id': 预测的类别ID, 'pred_category_name': 预测的类别名称 } """ best_pred = None best_iou = 0 for pred in predictions: iou = compute_iou(pred['bbox'], gt_bbox) if iou > best_iou: best_iou = iou best_pred = pred if best_pred is None or best_iou < min_iou: return { 'pred': None, 'iou': best_iou, 'localized': False, 'classified_correct': False, 'pred_category_id': None, 'pred_category_name': None } pred_cat_id = best_pred['category_id'] pred_cat_name = id_to_name.get(pred_cat_id, 'unknown') classified_correct = (pred_cat_id == gt_category_id) return { 'pred': best_pred, 'iou': best_iou, 'localized': True, 'classified_correct': classified_correct, 'pred_category_id': pred_cat_id, 'pred_category_name': pred_cat_name } def analyze_single_gt( gt_ann: dict, declip_preds: List[dict], clipself_preds: List[dict], clip_preds: List[dict], category_name: str, is_novel: bool, id_to_name: dict, iou_threshold: float = 0.5 ) -> dict: """ 分析单个GT目标在三个模型上的检测结果。 核心逻辑（针对OVD优化）： 1. 在所有预测中找IoU最高的框（不限类别） 2. 判断分类是否正确 3. 计算优势分数：DeCLIP分类正确 + 其他模型分类错误 """ gt_bbox = gt_ann['bbox'] gt_category_id = gt_ann['category_id'] gt_area = gt_ann['area'] size_category = get_size_category(gt_area) # 在所有预测中找IoU最高的框（不限类别） declip_result = find_best_matching_pred(declip_preds, gt_bbox, gt_category_id, id_to_name, min_iou=0.1) clipself_result = find_best_matching_pred(clipself_preds, gt_bbox, gt_category_id, id_to_name, min_iou=0.1) clip_result = find_best_matching_pred(clip_preds, gt_bbox, gt_category_id, id_to_name, min_iou=0.1) # 提取结果 declip_pred = declip_result['pred'] clipself_pred = clipself_result['pred'] clip_pred = clip_result['pred'] declip_iou = declip_result['iou'] clipself_iou = clipself_result['iou'] clip_iou = clip_result['iou'] # 定位成功 = IoU >= threshold declip_localized = declip_iou >= iou_threshold clipself_localized = clipself_iou >= iou_threshold clip_localized = clip_iou >= iou_threshold # 分类正确 = 定位成功 + 预测类别等于GT类别 declip_classified_correct = declip_localized and declip_result['classified_correct'] clipself_classified_correct = clipself_localized and clipself_result['classified_correct'] clip_classified_correct = clip_localized and clip_result['classified_correct'] # 计算DeCLIP优势分数（基于分类能力） advantage_score = 0 advantage_type = 'none' if declip_classified_correct: # DeCLIP分类正确 if not clipself_classified_correct and not clip_classified_correct: # 只有DeCLIP分类正确（最强优势） advantage_score = 3.0 + declip_iou advantage_type = 'unique_correct_classification' elif not clipself_classified_correct or not clip_classified_correct: # DeCLIP + 部分模型分类正确 advantage_score = 2.0 + declip_iou advantage_type = 'partial_correct_classification' else: # 三个模型都分类正确，比较IoU min_iou_advantage = min(declip_iou - clipself_iou, declip_iou - clip_iou) if min_iou_advantage > 0.05: advantage_score = 1.0 + min_iou_advantage advantage_type = 'better_iou' return { 'gt_id': gt_ann['id'], 'image_id': gt_ann['image_id'], 'category_id': gt_category_id, 'category_name': category_name, 'is_novel': is_novel, 'gt_bbox': gt_bbox, 'gt_area': gt_area, 'size_category': size_category, # DeCLIP结果 'declip_localized': declip_localized, 'declip_classified_correct': declip_classified_correct, 'declip_iou': declip_iou, 'declip_score': declip_pred['score'] if declip_pred else 0, 'declip_bbox': declip_pred['bbox'] if declip_pred else None, 'declip_pred_category': declip_result['pred_category_name'], # CLIPSelf结果 'clipself_localized': clipself_localized, 'clipself_classified_correct': clipself_classified_correct, 'clipself_iou': clipself_iou, 'clipself_score': clipself_pred['score'] if clipself_pred else 0, 'clipself_bbox': clipself_pred['bbox'] if clipself_pred else None, 'clipself_pred_category': clipself_result['pred_category_name'], # CLIP结果 'clip_localized': clip_localized, 'clip_classified_correct': clip_classified_correct, 'clip_iou': clip_iou, 'clip_score': clip_pred['score'] if clip_pred else 0, 'clip_bbox': clip_pred['bbox'] if clip_pred else None, 'clip_pred_category': clip_result['pred_category_name'], # 优势分析 'advantage_score': advantage_score, 'advantage_type': advantage_type, 'iou_advantage_vs_clipself': declip_iou - clipself_iou, 'iou_advantage_vs_clip': declip_iou - clip_iou, } def compute_statistics(results: List[dict], iou_threshold: float = 0.5) -> dict: """ 计算统计信息。 新增指标： - localized: 定位成功（IoU >= threshold） - classified_correct: 分类正确（定位成功 + 类别正确） """ stats = { 'total': len(results), 'iou_threshold': iou_threshold, 'by_size': {size: { 'total': 0, 'declip_localized': 0, 'clipself_localized': 0, 'clip_localized': 0, 'declip_correct': 0, 'clipself_correct': 0, 'clip_correct': 0, 'declip_iou_sum': 0, 'clipself_iou_sum': 0, 'clip_iou_sum': 0 } for size in ['small', 'medium', 'large']}, 'by_category_type': { 'novel': {'total': 0, 'declip_correct': 0, 'clipself_correct': 0, 'clip_correct': 0}, 'base': {'total': 0, 'declip_correct': 0, 'clipself_correct': 0, 'clip_correct': 0} }, 'by_size_and_type': {} } # 初始化by_size_and_type for size in ['small', 'medium', 'large']: for cat_type in ['novel', 'base']: key = f"{size}_{cat_type}" stats['by_size_and_type'][key] = { 'total': 0, 'declip_localized': 0, 'clipself_localized': 0, 'clip_localized': 0, 'declip_correct': 0, 'clipself_correct': 0, 'clip_correct': 0, 'declip_unique_correct': 0, # 只有DeCLIP分类正确 'declip_better_iou': 0, # DeCLIP IoU最高 'declip_iou_sum': 0, 'clipself_iou_sum': 0, 'clip_iou_sum': 0, } for r in results: size = r['size_category'] cat_type = 'novel' if r['is_novel'] else 'base' key = f"{size}_{cat_type}" # by_size stats['by_size'][size]['total'] += 1 if r['declip_localized']: stats['by_size'][size]['declip_localized'] += 1 stats['by_size'][size]['declip_iou_sum'] += r['declip_iou'] if r['clipself_localized']: stats['by_size'][size]['clipself_localized'] += 1 stats['by_size'][size]['clipself_iou_sum'] += r['clipself_iou'] if r['clip_localized']: stats['by_size'][size]['clip_localized'] += 1 stats['by_size'][size]['clip_iou_sum'] += r['clip_iou'] if r['declip_classified_correct']: stats['by_size'][size]['declip_correct'] += 1 if r['clipself_classified_correct']: stats['by_size'][size]['clipself_correct'] += 1 if r['clip_classified_correct']: stats['by_size'][size]['clip_correct'] += 1 # by_category_type stats['by_category_type'][cat_type]['total'] += 1 if r['declip_classified_correct']: stats['by_category_type'][cat_type]['declip_correct'] += 1 if r['clipself_classified_correct']: stats['by_category_type'][cat_type]['clipself_correct'] += 1 if r['clip_classified_correct']: stats['by_category_type'][cat_type]['clip_correct'] += 1 # by_size_and_type stats['by_size_and_type'][key]['total'] += 1 if r['declip_localized']: stats['by_size_and_type'][key]['declip_localized'] += 1 stats['by_size_and_type'][key]['declip_iou_sum'] += r['declip_iou'] if r['clipself_localized']: stats['by_size_and_type'][key]['clipself_localized'] += 1 stats['by_size_and_type'][key]['clipself_iou_sum'] += r['clipself_iou'] if r['clip_localized']: stats['by_size_and_type'][key]['clip_localized'] += 1 stats['by_size_and_type'][key]['clip_iou_sum'] += r['clip_iou'] if r['declip_classified_correct']: stats['by_size_and_type'][key]['declip_correct'] += 1 if r['clipself_classified_correct']: stats['by_size_and_type'][key]['clipself_correct'] += 1 if r['clip_classified_correct']: stats['by_size_and_type'][key]['clip_correct'] += 1 # DeCLIP优势统计 if r['declip_classified_correct'] and not r['clipself_classified_correct'] and not r['clip_classified_correct']: stats['by_size_and_type'][key]['declip_unique_correct'] += 1 if r['declip_iou'] > r['clipself_iou'] and r['declip_iou'] > r['clip_iou']: stats['by_size_and_type'][key]['declip_better_iou'] += 1 # 计算分类正确率（Localized + Correct）和定位率 for key in stats['by_size_and_type']: s = stats['by_size_and_type'][key] total = max(s['total'], 1) s['declip_cls_acc'] = s['declip_correct'] / total s['clipself_cls_acc'] = s['clipself_correct'] / total s['clip_cls_acc'] = s['clip_correct'] / total s['declip_loc_rate'] = s['declip_localized'] / total s['clipself_loc_rate'] = s['clipself_localized'] / total s['clip_loc_rate'] = s['clip_localized'] / total for size in stats['by_size']: s = stats['by_size'][size] total = max(s['total'], 1) s['declip_cls_acc'] = s['declip_correct'] / total s['clipself_cls_acc'] = s['clipself_correct'] / total s['clip_cls_acc'] = s['clip_correct'] / total return stats def main(): args = parse_args() print("=" * 70) print("OV-COCO Cross-Model Comparison Analysis V2") print("=" * 70) # 加载数据 print("\n[1/5] Loading data...") print(f" Loading annotations: {args.ann_file}") coco_gt = COCO(args.ann_file) print(f" Loading seen classes: {args.seen_classes}") with open(args.seen_classes, 'r') as f: seen_classes = set(json.load(f)) id_to_name = {cat['id']: cat['name'] for cat in coco_gt.dataset['categories']} print(f" Loading DeCLIP predictions...") declip_preds = load_predictions(args.declip_pred) print(f" Loading CLIPSelf predictions...") clipself_preds = load_predictions(args.clipself_pred) print(f" Loading CLIP predictions...") clip_preds = load_predictions(args.clip_pred) print(f"\n Total images: {len(coco_gt.imgs)}") print(f" Total GT annotations: {len(coco_gt.anns)}") # 分析每个GT目标 print("\n[2/5] Analyzing each GT object...") all_results = [] novel_results = [] for ann_id in coco_gt.anns: ann = coco_gt.anns[ann_id] img_id = ann['image_id'] category_name = id_to_name.get(ann['category_id'], 'unknown') is_novel = category_name not in seen_classes # 获取该图像的所有预测（不按类别分组） img_declip_preds = declip_preds.get(img_id, []) img_clipself_preds = clipself_preds.get(img_id, []) img_clip_preds = clip_preds.get(img_id, []) result = analyze_single_gt( ann, img_declip_preds, img_clipself_preds, img_clip_preds, category_name, is_novel, id_to_name, args.iou_threshold ) all_results.append(result) if is_novel: novel_results.append(result) print(f" Analyzed {len(all_results)} GT objects") print(f" Novel category objects: {len(novel_results)}") # 计算统计 print("\n[3/5] Computing statistics...") all_stats = compute_statistics(all_results, args.iou_threshold) novel_stats = compute_statistics(novel_results, args.iou_threshold) # 找出DeCLIP优势case（只保留 small + novel） # 优势定义：DeCLIP分类正确 + 其他模型分类错误 print("\n[4/5] Finding DeCLIP advantage cases...") advantage_cases = [ r for r in novel_results if r['advantage_score'] > 0 and r['size_category'] == 'small' ] # 按 advantage_score 降序排列（优先 unique_correct_classification），然后按 IoU 降序 advantage_cases = sorted(advantage_cases, key=lambda x: (x['advantage_score'], x['declip_iou']), reverse=True) top_cases = advantage_cases[:args.top_k] print(f" Found {len(advantage_cases)} small novel cases where DeCLIP has classification advantage") print(f" Top {args.top_k} cases selected") # 找出DeCLIP未解决的 novel 小目标 case（分类不正确） unsolved_cases = [ r for r in novel_results if r['size_category'] == 'small' and not r['declip_classified_correct'] ] # 优先展示定位更接近但分类错误的case（IoU从大到小） unsolved_cases = sorted(unsolved_cases, key=lambda x: x['declip_iou'], reverse=True)[:args.top_k] total_unsolved = len([r for r in novel_results if r['size_category'] == 'small' and not r['declip_classified_correct']]) print(f" Found {total_unsolved} unsolved small novel cases (DeCLIP classification incorrect)") print(f" Top {args.top_k} unsolved cases selected") # 保存结果 print("\n[5/5] Saving results...") output_dir = Path(args.output) output_dir.mkdir(parents=True, exist_ok=True) with open(output_dir / 'all_results.json', 'w') as f: json.dump(all_results, f, indent=2) with open(output_dir / 'novel_results.json', 'w') as f: json.dump(novel_results, f, indent=2) with open(output_dir / 'top_advantage_cases.json', 'w') as f: json.dump(top_cases, f, indent=2) with open(output_dir / 'unsolved_small_novel_cases.json', 'w') as f: json.dump(unsolved_cases, f, indent=2) # 统计信息 top_by_size = defaultdict(list) for case in top_cases: top_by_size[case['size_category']].append(case) stats_output = { 'config': { 'declip_pred': args.declip_pred, 'clipself_pred': args.clipself_pred, 'clip_pred': args.clip_pred, 'iou_threshold': args.iou_threshold, }, 'all_categories': all_stats, 'novel_categories': novel_stats, 'top_cases_summary': { 'total': len(top_cases), 'by_size': {size: len(cases) for size, cases in top_by_size.items()}, 'advantage_types': { 'unique_correct_classification': sum(1 for c in top_cases if c['advantage_type'] == 'unique_correct_classification'), 'partial_correct_classification': sum(1 for c in top_cases if c['advantage_type'] == 'partial_correct_classification'), 'better_iou': sum(1 for c in top_cases if c['advantage_type'] == 'better_iou') } } } with open(output_dir / 'statistics.json', 'w') as f: json.dump(stats_output, f, indent=2) # 打印关键统计 print("\n" + "=" * 70) print("SUMMARY - Novel Categories (OVD Analysis)") print("=" * 70) print("\n[Classification Accuracy by Size] (Localized + Correct Category)") print("-" * 75) print(f"{'Size':<10} {'Total':<8} {'DeCLIP':<18} {'CLIPSelf':<15} {'CLIP':<15}") print("-" * 75) for size in ['small', 'medium', 'large']: key = f"{size}_novel" s = novel_stats['by_size_and_type'].get(key, {}) total = s.get('total', 0) if total > 0: declip_acc = s.get('declip_cls_acc', 0) * 100 clipself_acc = s.get('clipself_cls_acc', 0) * 100 clip_acc = s.get('clip_cls_acc', 0) * 100 delta = declip_acc - clipself_acc print(f"{size:<10} {total:<8} {declip_acc:>6.1f}% (+{delta:>4.1f}) {clipself_acc:>6.1f}% {clip_acc:>6.1f}%") print("\n[DeCLIP Classification Advantages]") print("-" * 75) print(f"{'Size':<10} {'Unique Correct':<18} {'Better IoU':<15} {'Total Adv.':<15}") print("-" * 75) for size in ['small', 'medium', 'large']: key = f"{size}_novel" s = novel_stats['by_size_and_type'].get(key, {}) unique_correct = s.get('declip_unique_correct', 0) better = s.get('declip_better_iou', 0) total_adv = sum(1 for c in advantage_cases if c['size_category'] == size) print(f"{size:<10} {unique_correct:<18} {better:<15} {total_adv:<15}") print(f"\nResults saved to: {output_dir}") print("=" * 70) if __name__ == '__main__': main()