evaluation.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

"""
-------------------------------------------------
   Author:       Kuihao Wang
   Email:        wangkuihao2025@mails.szu.edu.cn
   Date:         2026/03/13
-------------------------------------------------
"""

"""
COCO Metrics Evaluator

This is an independent script for calculating standard COCO evaluation metrics (mAP, AP50, AR, etc.).
The script simultaneously outputs evaluation results for both Instance Segmentation and Bounding Box.
Ideal for evaluating models that output polygons (e.g., building footprint extraction, 
remote sensing instance segmentation, etc.).

[Dependencies]
pip install pycocotools

[Usage]
python eval_coco.py --gt path/to/ground_truth.json --dt path/to/predictions.json

[Input File Requirements]
1. Ground Truth (--gt):
   Must be a standard COCO dataset JSON dictionary containing 'images', 'annotations', and 'categories'.
   For polygon annotations, the 'segmentation' field should be a nested list:
   "segmentation": [[x1, y1, x2, y2, x3, y3, ...]]

2. Predictions (--dt):
   Must be a JSON list containing prediction results.
   ⚠️ Required fields: 'image_id', 'category_id', 'bbox', 'segmentation', 'score'
   Example:
   [
     {
       "image_id": 1,
       "category_id": 1,
       "bbox": [100.0, 150.0, 50.0, 60.0],
       "segmentation": [[100.0, 150.0, 150.0, 150.0, 150.0, 210.0, 100.0, 210.0]],
       "score": 0.952   // Confidence score; COCOeval relies on this for ranking to calculate mAP
     }
   ]

[Output Description]
The script prints two official COCO format evaluation tables to the terminal:
- AP @ IoU=0.50:0.95 (Primary metric: mAP)
- AP @ IoU=0.50 (AP50)
- AP @ IoU=0.75 (AP75)
- AP for small/medium/large objects (Accuracy across different scales)
- AR (Average Recall)
"""

import argparse
import os
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval

def main():
    # Initialize command-line argument parser
    parser = argparse.ArgumentParser(
        description="Calculate standard COCO metrics (BBox & Segmentation) for model predictions."
    )
    parser.add_argument(
        "--gt", 
        type=str, 
        required=True, 
        help="Path to Ground Truth JSON file (standard COCO annotation format)."
    )
    parser.add_argument(
        "--dt", 
        type=str, 
        required=True, 
        help="Path to Detection/Prediction JSON file (list of dicts with 'score')."
    )
    args = parser.parse_args()

    # Check if files exist to prevent errors
    if not os.path.exists(args.gt):
        raise FileNotFoundError(f"Ground Truth file not found: {args.gt}")
    if not os.path.exists(args.dt):
        raise FileNotFoundError(f"Prediction file not found: {args.dt}")

    # ==========================================
    # 1. Load Ground Truth
    # ==========================================
    print(f"[*] Loading Ground Truth from: {args.gt}")
    # The COCO class automatically parses the JSON and indexes images, categories, and annotations
    cocoGt = COCO(args.gt)

    # ==========================================
    # 2. Load Predictions
    # ==========================================
    print(f"[*] Loading Predictions from: {args.dt}")
    # loadRes creates a new COCO object for evaluation based on the prediction list and GT structure
    cocoDt = cocoGt.loadRes(args.dt)

    # ==========================================
    # 3. Execute Evaluation: Instance Segmentation (Mask/Polygons)
    # ==========================================
    print("\n" + "="*50)
    print("Running COCO Evaluation (Segmentation Mask/Polygons)")
    print("="*50)
    # Instantiate COCOeval, specifying 'segm' to perform IoU calculations on the segmentation field
    cocoEval_segm = COCOeval(cocoGt, cocoDt, 'segm')
    cocoEval_segm.evaluate()    # Matches detections to ground truth per image/category
    cocoEval_segm.accumulate()  # Accumulates evaluation results across all images
    cocoEval_segm.summarize()   # Prints the summary metrics table (mAP, AR, etc.)

    # ==========================================
    # 4. Execute Evaluation: Object Detection (Bounding Box)
    # ==========================================
    print("\n" + "="*50)
    print("Running COCO Evaluation (Bounding Box)")
    print("="*50)
    # Instantiate COCOeval, specifying 'bbox' to perform IoU calculations on the bbox field
    cocoEval_bbox = COCOeval(cocoGt, cocoDt, 'bbox')
    cocoEval_bbox.evaluate()
    cocoEval_bbox.accumulate()
    cocoEval_bbox.summarize()

if __name__ == "__main__":
    main()