evaluation.py

import os
import json
import cv2
import glob
import numpy as np
from tqdm import tqdm
from time import time
from sklearn.metrics import jaccard_score, precision_score
from concurrent.futures import ThreadPoolExecutor, as_completed

def mask_to_bbox(mask):
    # mask: 2D np.uint8 array, 0/1
    idx = np.where(mask > 0)
    if len(idx[0]) == 0:
        return None
    y1, x1 = idx[0].min(), idx[1].min()
    y2, x2 = idx[0].max(), idx[1].max()
    return x1, y1, x2, y2

def compute_giou(pred, label):
    '''
    pred: 2D np.uint8 array, 0/1
    label: 2D np.uint8 array, 0/1
    return: giou score
    '''
    bbox1 = mask_to_bbox(pred)
    bbox2 = mask_to_bbox(label)
    if bbox1 is None or bbox2 is None:
        return 0.0
    xA = max(bbox1[0], bbox2[0])
    yA = max(bbox1[1], bbox2[1])
    xB = min(bbox1[2], bbox2[2])
    yB = min(bbox1[3], bbox2[3])
    interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
    boxAArea = (bbox1[2] - bbox1[0] + 1) * (bbox1[3] - bbox1[1] + 1)
    boxBArea = (bbox2[2] - bbox2[0] + 1) * (bbox2[3] - bbox2[1] + 1)
    unionArea = boxAArea + boxBArea - interArea
    iou = interArea / unionArea if unionArea > 0 else 0.0
    xC = min(bbox1[0], bbox2[0])
    yC = min(bbox1[1], bbox2[1])
    xD = max(bbox1[2], bbox2[2])
    yD = max(bbox1[3], bbox2[3])
    encloseArea = (xD - xC + 1) * (yD - yC + 1)
    giou = iou - (encloseArea - unionArea) / encloseArea if encloseArea > 0 else iou
    return giou

def compute_ciou(pred, label):
    '''
    pred: 2D np.uint8 array, 0/1
    label: 2D np.uint8 array, 0/1
    return: ciou score
    '''
    bbox1 = mask_to_bbox(pred)
    bbox2 = mask_to_bbox(label)
    if bbox1 is None or bbox2 is None:
        return 0.0
    xA = max(bbox1[0], bbox2[0])
    yA = max(bbox1[1], bbox2[1])
    xB = min(bbox1[2], bbox2[2])
    yB = min(bbox1[3], bbox2[3])
    interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)
    boxAArea = (bbox1[2] - bbox1[0] + 1) * (bbox1[3] - bbox1[1] + 1)
    boxBArea = (bbox2[2] - bbox2[0] + 1) * (bbox2[3] - bbox2[1] + 1)
    unionArea = boxAArea + boxBArea - interArea
    iou = interArea / unionArea if unionArea > 0 else 0.0
    c1x = (bbox1[0] + bbox1[2]) / 2
    c1y = (bbox1[1] + bbox1[3]) / 2
    c2x = (bbox2[0] + bbox2[2]) / 2
    c2y = (bbox2[1] + bbox2[3]) / 2
    center_dist = (c1x - c2x) ** 2 + (c1y - c2y) ** 2
    xC = min(bbox1[0], bbox2[0])
    yC = min(bbox1[1], bbox2[1])
    xD = max(bbox1[2], bbox2[2])
    yD = max(bbox1[3], bbox2[3])
    diag = (xD - xC) ** 2 + (yD - yC) ** 2
    ciou = iou - center_dist / diag if diag > 0 else iou
    return ciou


def to_binary_mask(img, flip=False):
    """自动将0/1或0/255掩码转为0/1二值掩码"""
    if flip:
        if img.max() == 1:
            return (1 - img).astype(np.uint8)
        elif img.max() > 1:
            return (img <= 127).astype(np.uint8)
        else:
            return np.zeros_like(img, dtype=np.uint8)
    else:
        if img.max() == 1:
            return img.astype(np.uint8)
        elif img.max() > 1:
            return (img > 127).astype(np.uint8)
        else:
            return np.zeros_like(img, dtype=np.uint8)

def process_sample(sample, result_dir, debug=False, dataset='GeoPixInstruct'):
    # base_name = os.path.splitext(os.path.basename(sample["label_path"]))[0]
    base_name = os.path.splitext(os.path.basename(sample["image_path"]))[0]
    try:
        if debug:
            print(f"[DEBUG] Processing sample: {base_name}")
        possible_files = glob.glob(os.path.join(result_dir, base_name + "*.*"))
        if not possible_files:
            print(f"[WARN] Prediction file not found for {base_name} in {result_dir}, all metrics set to 0")
            return (0.0, 0.0, 0.0, [0.0, 0.0, 0.0, 0.0, 0.0])
        pred_path = possible_files[0]
        label_path = sample["label_path"]
        
        if dataset == "GeoPixInstruct":
            label_path = label_path.split('Data-Source/')[-1][:-6]
            if label_path.endswith('_'):
                label_path = label_path[:-1]
            label_path = '/home/l/Files/' + label_path.replace('labels_valid', 'gray') + '.png'
        elif dataset == "EarthReason":
            label_path = '/home/l/YanJiafeng/GRASP/' + label_path.split('RS-RS/')[-1]
        elif dataset == "GRASP":
            label_path = '/home/l/Files/GRASP_test/' + label_path

        if debug:
            print(f"[DEBUG] label_path: {label_path}, pred_path: {pred_path}")
        label = cv2.imread(label_path, cv2.IMREAD_GRAYSCALE)
        pred = cv2.imread(pred_path, cv2.IMREAD_GRAYSCALE)
        if label is None:
            print(f"[ERROR] Failed to read label: {label_path}")
            return (0.0, 0.0, 0.0, [0.0, 0.0, 0.0, 0.0, 0.0])
        if pred is None:
            print(f"[ERROR] Failed to read pred: {pred_path}")
            return (0.0, 0.0, 0.0, [0.0, 0.0, 0.0, 0.0, 0.0])

        # 关键：自动兼容0/1和0/255
        label_bin = to_binary_mask(label, flip=(dataset=="GeoPixInstruct"))
        pred_bin = to_binary_mask(pred)

        if label_bin.shape != pred_bin.shape:
            print(f"[ERROR] Shape mismatch: label {label_bin.shape}, pred {pred_bin.shape} for {base_name}")
            return (0.0, 0.0, 0.0, [0.0, 0.0, 0.0, 0.0, 0.0])
        iou = jaccard_score(label_bin.flatten(), pred_bin.flatten(), zero_division=0)
        giou = compute_giou(pred_bin, label_bin)
        ciou = compute_ciou(pred_bin, label_bin)
        precisions = []
        for t in [0.5]:
            p = precision_score(label_bin.flatten(), pred_bin.flatten(), zero_division=0)
            precisions.append(p)
        if debug:
            print(f"[DEBUG] {base_name} done: IoU={iou:.4f}, gIoU={giou:.4f}, cIoU={ciou:.4f}")
        return (iou, giou, ciou, precisions)
    except Exception as e:
        print(f"[EXCEPTION] Error processing {base_name}: {e}")
        import traceback
        traceback.print_exc()
        return (0.0, 0.0, 0.0, [0.0, 0.0, 0.0, 0.0, 0.0])
    

import pandas as pd
import os
from openpyxl import load_workbook

def save_metrics(data_dict, file_path="metrics.xlsx", sheet_name="Sheet1", mode="append"): 
    """
    将字典增量保存到 Excel 文件
    
    参数:
        data_dict (dict): 要保存的字典数据
        file_path (str): Excel 文件路径
        sheet_name (str): 工作表名称 (默认 "Sheet1")
        mode (str): 保存模式 - "append" (追加) 或 "replace" (替换整个工作表)
    """
    # 检查文件是否存在
    file_exists = os.path.isfile(file_path)
    
    # 如果文件不存在或选择替换模式，直接创建新文件
    if not file_exists or mode == "replace":
        df = pd.DataFrame([data_dict])
        df.to_excel(file_path, sheet_name=sheet_name, index=False)
        print(f"{'创建' if not file_exists else '替换'}Excel文件: {file_path}")
        return
    
    # 追加模式 - 文件存在
    try:
        # 读取现有数据
        with pd.ExcelFile(file_path) as xls:
            if sheet_name in xls.sheet_names:
                existing_df = pd.read_excel(xls, sheet_name=sheet_name)
            else:
                existing_df = pd.DataFrame()
        
        # 创建新数据DataFrame
        new_df = pd.DataFrame([data_dict])
        
        # 合并数据
        combined_df = pd.concat([existing_df, new_df], ignore_index=True)
        
        # 保存回Excel（保留其他工作表）
        with pd.ExcelWriter(file_path, engine='openpyxl', mode='a', if_sheet_exists='replace') as writer:
            combined_df.to_excel(writer, sheet_name=sheet_name, index=False)
        
        print(f"成功使用pandas追加数据到Excel文件: {file_path}")
    
    except Exception as e2:
        print(f"所有方法失败: {e2}")
        raise RuntimeError("无法保存数据到Excel文件")
    
def evaluate_segmentation(qa_json_path, result_dir, num_workers=8, debug=False, dataset='GeoPixInstruct'):
    with open(qa_json_path, 'r', encoding='utf-8') as f:
        samples = json.load(f)

    metrics = {
        "mIoU": [],
        "gIoU": [],
        "cIoU": [],
        "precision@0.5": []
    }
    print(f'Currently evaluating: {dataset}')

    start_time = time()
    results = []
    with ThreadPoolExecutor(max_workers=num_workers) as executor:
        futures = [executor.submit(process_sample, sample, result_dir, debug, dataset) for sample in samples]
        for idx, f in enumerate(tqdm(as_completed(futures), total=len(futures), desc="Evaluating")):
            try:
                res = f.result()
                if res is not None:
                    iou, giou, ciou, precisions = res
                    metrics["mIoU"].append(iou)
                    metrics["gIoU"].append(giou)
                    metrics["cIoU"].append(ciou)
                    for pidx, t in enumerate([0.5]):
                        metrics[f"precision@{t}"].append(precisions[pidx])
                else:
                    if debug:
                        print(f"[DEBUG] Future {idx} returned None")
            except Exception as e:
                print(f"[EXCEPTION] Error in future {idx}: {e}")
                import traceback
                traceback.print_exc()

    elapsed = time() - start_time
    print(f"\n✅ Evaluation completed in {elapsed:.2f} seconds")

    avg_metrics = {k: float(np.mean(v)) if v else 0.0 for k, v in metrics.items()}
    print("\n📊 Average Metrics:")
    for k, v in avg_metrics.items():
        print(f"{k}: {v:.4f}")
    
    model_dataset_name = result_dir.split("contrast_test_result/")[-1].split("/Binary")[0]
    model_name, dataset_name = model_dataset_name.split("/")
    model_size = result_dir.split('-')[-1]
    if len(model_size) <= 3:
        model_name += '-' + model_size
    avg_metrics['model_name'] = model_name
    avg_metrics['dataset_name'] = dataset_name

    
    save_metrics(avg_metrics)

    return avg_metrics

# Example usage:
if __name__ == "__main__":
    # ### GeoPixInstruct
    # # qa_json = "/home/l/YanJiafeng/GRASP/GRASP-data/in_domain_test/GeoPixInstruct/test.json"
    # ### EarthReason
    # qa_json = "/home/l/YanJiafeng/GRASP/GRASP-data/in_domain_test/EarthReason/test.json"

    # ### GeoPixInstruct
    # # result_folder = "/home/l/Files/contrast_test_result/GeoPix/GeoPixInstruct/Overlay"
    # ### EarthReason
    # result_folder = "/home/l/Files/contrast_test_result/GeoPix/EarthReason/Binary"
    # evaluate_segmentation(qa_json, result_folder, num_workers=16)

    qa_json = {
        "GeoPixInstruct": "/home/l/YanJiafeng/GRASP/GRASP-data/in_domain_test/GeoPixInstruct/test.json",
        "EarthReason": "/home/l/YanJiafeng/GRASP/GRASP-data/in_domain_test/EarthReason/test.json",
        "GRASP": "/home/l/Files/GRASP_test/grasp_test_annotation.json"
    }
    result_folder = {
        "GeoPixInstruct": [
            # '/home/l/Files/contrast_test_result/GeoPix/GeoPixInstruct/Binary',
            '/home/l/Files/GRASP_contrast_results/contrast_test_result/GeoPixel/GeoPixInstruct/Binary',
            # '/home/l/Files/contrast_test_result/LISA/GeoPixInstruct/Binary-7b',
            # '/home/l/Files/contrast_test_result/LISA/GeoPixInstruct/Binary-13b',
            # '/home/l/Files/contrast_test_result/PixelLM/GeoPixInstruct/Binary-7b',
            # '/home/l/Files/contrast_test_result/PixelLM/GeoPixInstruct/Binary-13b',
            # '/home/l/Files/contrast_test_result/Sa2VA/GeoPixInstruct/Binary-8b',
            # '/home/l/Files/contrast_test_result/SegEarth-R1/GeoPixInstruct/Binary',

            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/RMSIN/GeoPixInstruct/Binary',
            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/rrsis/GeoPixInstruct/Binary',
            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/RSRefSeg2/GeoPixInstruct/Binary',
        ],
        "EarthReason": [
            # '/home/l/Files/contrast_test_result/GeoPix/EarthReason/Binary',
            # '/home/l/Files/contrast_test_result/GeoPixel/EarthReason/Binary',
            # '/home/l/Files/contrast_test_result/LISA/EarthReason/Binary-7b',
            # '/home/l/Files/contrast_test_result/LISA/EarthReason/Binary-13b',
            # '/home/l/Files/contrast_test_result/PixelLM/EarthReason/Binary-7b',
            # '/home/l/Files/contrast_test_result/PixelLM/EarthReason/Binary-13b',
            # '/home/l/Files/contrast_test_result/Sa2VA/EarthReason/Binary-8b',
            # '/home/l/Files/contrast_test_result/SegEarth-R1/EarthReason/Binary',

            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/RMSIN/EarthReason/Binary',
            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/rrsis/EarthReason/Binary',
            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/RSRefSeg2/EarthReason/Binary',
        ],
        "GRASP": [
            # '/home/l/Files/contrast_test_result/GeoPix/GRASP/Binary',
            # '/home/l/Files/contrast_test_result/GeoPixel/GRASP/Binary',
            # '/home/l/Files/contrast_test_result/LISA/GRASP/Binary-7b',
            # '/home/l/Files/contrast_test_result/PixelLM/GRASP/Binary-7b',
            # '/home/l/Files/contrast_test_result/Sa2VA/GRASP/Binary-8b',
            # '/home/l/Files/contrast_test_result/SegEarth-R1/GRASP/Binary',

            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/RMSIN/GRASP/Binary',
            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/rrsis/GRASP/Binary',
            # '/home/l/Files/GRASP_contrast_results/contrast_test_result/RSRefSeg2/GRASP/Binary',
        ]
    }
    debug = False

    for dataset in ['GeoPixInstruct', 'EarthReason', 'GRASP']:
        qa = qa_json[dataset]
        for result_dir in result_folder[dataset]:
            evaluate_segmentation(qa, result_dir, num_workers=16, debug=debug, dataset=dataset)