pixel accuracy and IoU for semantic segmentation

ss_metrics.py

+# -*- coding: utf-8 -*-
+
+import sys
+import numpy as np
+from PIL import Image
+
+# CLASS = 14
+CLASS = 2
+
+def main():
+  
+  img = Image.open(sys.argv[1])
+  w, h = img.size
+  target = np.asarray(img, dtype=np.int64)
+  # print(target.shape)
+
+  img = Image.open(sys.argv[2])
+  pred = np.asarray(img, dtype=np.int64)[:,:,0]
+  # print(pred.shape)
+
+  acc = np.zeros(CLASS, dtype=np.int64)    # iをiと判定した数 acc[i]
+  nacc = np.zeros(CLASS, dtype=np.int64)   # iをi以外と判定した数 nacc[i]
+  occ = np.zeros(CLASS, dtype=np.int64)    # iに本来所属する画素数 occ[i]
+  pcc = np.zeros(CLASS, dtype=np.int64)    # iに所属するとされた画素数 pcc[i]
+
+  accuracy = np.zeros(CLASS, dtype=np.float64)
+  iou = np.zeros(CLASS, dtype=np.float64)
+
+  print(sys.argv[1])
+  
+  for y in range(0, h):
+    for x in range(0, w):
+      occ[target[y][x]] += 1
+      pcc[pred[y][x]] += 1      
+      
+      if target[y][x] == pred[y][x]:
+        acc[target[y][x]] += 1
+      else:
+        nacc[target[y][x]] += 1
+
+  for i in range(0, CLASS):
+    if acc[i] + nacc[i] > 0:
+      print("Class {0:2d}: correct {1:6d}, incorrect {2:6d}, accuracy {3:.3}".format(i, acc[i], nacc[i], float(acc[i])/float(acc[i]+nacc[i])))
+      accuracy[i] = float(acc[i])/float(acc[i]+nacc[i])
+    else:
+      print("Class {0:2d}: correct      0, incorrect      0, accuracy     -".format(i))
+      accuracy[i] = -1
+
+  print("Total   : correct {0:6d}, incorrect {1:6d}, accuracy {2:.3}".format(np.sum(acc), np.sum(nacc), float(np.sum(acc))/float(np.sum(acc)+np.sum(nacc))))
+
+  cc = 0
+  s = 0
+  for a in accuracy:
+    if a != -1:
+      cc += 1
+      s += a
+  print("Mean Accuracy   : {0:.3}".format(s/float(cc)))
+  print()
+
+  for i in range(0, CLASS):
+    if occ[i]+pcc[i]-acc[i] > 0:
+      iou[i] = acc[i] / (occ[i]+pcc[i]-acc[i])
+      print("Class {0:2d}: IoU {1:.3}".format(i, acc[i] / (occ[i]+pcc[i]-acc[i])))
+    else:
+      iou[i] = -1
+      print("Class {0:2d}: IoU     -".format(i))
+      
+  print("Total IoU  : {0:.3}".format(np.sum(acc) / float(np.sum(occ)+np.sum(pcc)-np.sum(acc))))
+  
+  cc = 0
+  s = 0
+  for a in iou:
+    if a != -1:
+      cc += 1
+      s += a
+  print("Mean IoU   : {0:.3}".format(s/float(cc)))
+  print()
+  
+if __name__ == "__main__":
+  main()