🐞 fix(Smeasure中的质心计算): 原始实现在输入过大时会溢出

原本直接基于np.sum()的实现在输入尺寸过大的时候会出现数值溢出。现在基于np.argwhere()的实现方式则避免了这一问题。

关于质心计算的更多细节可见文档：https://www.yuque.com/lart/blog/gpbigm

Author: lartpang

examples/test_metrics.py

@@ -4,11 +4,12 @@
 # @GitHub  : https://github.com/lartpang
 
 import os
+import sys
+from pprint import pprint
 
 import cv2
-from tqdm import tqdm
 
-# pip install pysodmetrics
+sys.path.append("..")
 from py_sod_metrics import MAE, Emeasure, Fmeasure, Smeasure, WeightedFmeasure
 
 FM = Fmeasure()
@@ -21,7 +22,8 @@
 mask_root = os.path.join(data_root, "masks")
 pred_root = os.path.join(data_root, "preds")
 mask_name_list = sorted(os.listdir(mask_root))
-for mask_name in tqdm(mask_name_list, total=len(mask_name_list)):
+for i, mask_name in enumerate(mask_name_list):
+    print(f"[{i}] Processing {mask_name}...")
     mask_path = os.path.join(mask_root, mask_name)
     pred_path = os.path.join(pred_root, mask_name)
     mask = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
@@ -50,24 +52,30 @@
     "maxFm": fm["curve"].max(),
 }
 
-print(results)
-# 'Smeasure': 0.9029763868504661,
-# 'wFmeasure': 0.5579812753638986,
-# 'MAE': 0.03705558476661653,
-# 'adpEm': 0.9408760066970631,
-# 'meanEm': 0.9566258293508715,
-# 'maxEm': 0.966954482892271,
-# 'adpFm': 0.5816750824038355,
-# 'meanFm': 0.577051059518767,
-# 'maxFm': 0.5886784581120638
+default_results = {
+    "v1_2_3": {
+        "Smeasure": 0.9029763868504661,
+        "wFmeasure": 0.5579812753638986,
+        "MAE": 0.03705558476661653,
+        "adpEm": 0.9408760066970631,
+        "meanEm": 0.9566258293508715,
+        "maxEm": 0.966954482892271,
+        "adpFm": 0.5816750824038355,
+        "meanFm": 0.577051059518767,
+        "maxFm": 0.5886784581120638,
+    },
+    "v1_3_0": {
+        "Smeasure": 0.9029761578759272,
+        "wFmeasure": 0.5579812753638986,
+        "MAE": 0.03705558476661653,
+        "adpEm": 0.9408760066970617,
+        "meanEm": 0.9566258293508704,
+        "maxEm": 0.9669544828922699,
+        "adpFm": 0.5816750824038355,
+        "meanFm": 0.577051059518767,
+        "maxFm": 0.5886784581120638,
+    },
+}
 
-# version 1.2.3
-# 'Smeasure': 0.9029763868504661,
-# 'wFmeasure': 0.5579812753638986,
-# 'MAE': 0.03705558476661653,
-# 'adpEm': 0.9408760066970631,
-# 'meanEm': 0.9566258293508715,
-# 'maxEm': 0.966954482892271,
-# 'adpFm': 0.5816750824038355,
-# 'meanFm': 0.577051059518767,
-# 'maxFm': 0.5886784581120638
+pprint(results)
+pprint({k: default_value - results[k] for k, default_value in default_results["v1_3_0"].items()})

py_sod_metrics/sod_metrics.py

@@ -266,19 +266,17 @@ def centroid(self, matrix: np.ndarray) -> tuple:
         so there is no need to use the redundant addition operation when dividing the region later,
         because the sequence generated by ``1:X`` in matlab will contain ``X``.
 
-        :param matrix: a data array
+        :param matrix: a bool data array
         :return: the centroid coordinate
         """
         h, w = matrix.shape
-        if matrix.sum() == 0:
+        area_object = np.count_nonzero(matrix)
+        if area_object == 0:
             x = np.round(w / 2)
             y = np.round(h / 2)
         else:
-            area_object = np.sum(matrix)
-            row_ids = np.arange(h)
-            col_ids = np.arange(w)
-            x = np.round(np.sum(np.sum(matrix, axis=0) * col_ids) / area_object)
-            y = np.round(np.sum(np.sum(matrix, axis=1) * row_ids) / area_object)
+            # More details can be found at: https://www.yuque.com/lart/blog/gpbigm
+            y, x = np.argwhere(matrix).mean(axis=0).round()
         return int(x) + 1, int(y) + 1
 
     def divide_with_xy(self, pred: np.ndarray, gt: np.ndarray, x: int, y: int) -> dict:

version.txt

@@ -1 +1 @@
-1.3.0
+1.3.1