[+] Add selective search process & scale factor threshold (computed from statistics and q(0.05)

data/train/predicted_labels/results_ss.csv

+0004, 515,801,541,851,0.9952337438579902,feux
+0004, 337,552,503,867,0.9894110951509656,feux
+0004, 407,326,424,364,0.9761382228356837,feux
+0004, 354,334,364,382,0.9752702712687679,feux
+0004, 425,602,458,648,0.9648697466863115,feux
+0004, 406,791,450,815,0.7745253440252448,obligation
+0004, 270,991,309,1000,0.7389454794497546,danger
+0039, 155,326,394,566,0.9999999972515112,ceder
+0039, 425,554,442,619,0.9999989042615994,feux
+0039, 194,755,233,779,0.9999988602516906,feux
+0039, 330,545,351,593,0.9937075776361155,feux
+0039, 640,770,715,790,0.9933633032725986,feux
+0039, 0,621,327,1000,0.9912733877457484,feux
+0039, 496,500,641,650,0.9870504749448945,feux
+0039, 246,536,284,568,0.9856224914127107,interdiction
+0039, 281,621,294,642,0.9785043961881216,obligation
+0039, 638,601,659,703,0.9738275396565923,feux
+0039, 693,534,715,572,0.9721666109721353,feux
+0039, 3,423,34,469,0.9613660899693652,feux
+0039, 255,485,277,508,0.9554668828090723,ceder
+0039, 158,751,191,770,0.9554464109950297,feux
+0039, 441,558,510,675,0.9542978068908017,feux
+0039, 368,572,405,605,0.7328675389698446,stop
+0039, 248,592,274,614,0.7253960393445719,ceder

supervised_learning/detection_selective_search.py

@@ -71,58 +71,67 @@ with open(output_file, "w") as f:
         print(f"[DETECTION] Processing image {name}")
 
         # 1. Extract rois from images with dynamic sliding window process
-        rois = []
-
 
-        rois = selective_search(image)
-        for roi in rois[:200]:
+        rois = []
+        rois_ss = selective_search(image)
+        for roi in rois_ss[:500]:
             x0, y0, x1, y1 = roi
-            window = np.array(Image.fromarray(image[y0:y1, x0:x1]).resize(AVERAGE_SIZE))
 
-            # HOG features
-            hog_features = np.array(hog(rgb2gray(window), pixels_per_cell=(16, 16), cells_per_block=(2, 2), block_norm='L2-Hys')).flatten()
+            area = (int(x1) - int(x0) + 1) * (int(y1) - int(y0) + 1)
             
-            # HUE features            
-            color_features = np.histogram(rgb2hsv(window)[:,:,0], bins=10, range=(0, 1), density=True)[0]
+            if area > AREA_THRESHOLD:
+                window = np.array(Image.fromarray(image[y0:y1, x0:x1]).resize(AVERAGE_SIZE))
+
+                # HOG features
+                hog_features = np.array(hog(rgb2gray(window), pixels_per_cell=(16, 16), cells_per_block=(2, 2), block_norm='L2-Hys')).flatten()
                 
-            # Concatenate ROI features
-            roi_features = np.concatenate((hog_features, color_features)).reshape(1, -1)
+                # HUE features            
+                color_features = np.histogram(rgb2hsv(window)[:,:,0], bins=10, range=(0, 1), density=True)[0]
+                    
+                # Concatenate ROI features
+                roi_features = np.concatenate((hog_features, color_features)).reshape(1, -1)
+                    
+                probas = {
+                    "danger" : None, 
+                    "interdiction": None,
+                    "obligation": None, 
+                    "stop": None,
+                    "ceder": None, 
+                }
+
+                for classe, classifier in classifiers.items():
+                    if classe not in ["feux", "fvert", "frouge", "forange"]:
+                        proba = classifier.predict_proba(roi_features)[0][1]
+                        if proba > 0.7:
+                            probas[classe] = proba
+                        else:
+                            probas[classe] = 0
                 
-            probas = {
-                "danger" : None, 
-                "interdiction": None,
-                "obligation": None, 
-                "stop": None,
-                "ceder": None, 
-            }
-
-            for classe, classifier in classifiers.items():
-                if classe not in ["feux", "fvert", "frouge", "forange"]:
-                    proba = classifier.predict_proba(roi_features)[0][1]
-                    if proba > 0.7:
-                        probas[classe] = proba
-                    else:
-                        probas[classe] = 0
-            
-            max_proba = 0
-            max_classe = "empty"
-            for classe, proba in probas.items():
-                if proba > max_proba:
+                max_proba = 0
+                max_classe = "empty"
+                for classe, proba in probas.items():
+                    if proba > max_proba:
+                        max_proba = proba
+                        max_classe = classe
+                
+                # Classificate if there is a light
+                proba = classifiers["feux"].predict_proba(roi_features)[0][1]
+                if proba > 0.95 and proba > max_proba:
                     max_proba = proba
-                    max_classe = classe
-            
-            # Classificate if there is a light
-            proba = classifiers["feux"].predict_proba(roi_features)[0][1]
-            if proba > 0.95 and proba > max_proba:
-                max_proba = proba
-                max_classe = "feux"
+                    max_classe = "feux"
 
-            if max_classe != "empty":
-                rois.append([x0, y0, x1, y1, max_classe, max_proba])             
+                if max_classe != "empty":
+                    rois.append([x0, y0, x1, y1, max_classe, max_proba])             
 
         # 2. Filter rois with Non Maximum Suppression process
         rois = non_max_suppression(rois, iou_threshold=0.1)     
-        #display_rois(image, rois)
+
+        for roi in rois:
+            x0, y0, x1, y1 = roi[:4]
+            area = (int(x1) - int(x0) + 1) * (int(y1) - int(y0) + 1)
+            print(area)
+
+        display_rois(image, rois)
 
         # 3. Write predicted labels into prediction files
         for roi in rois:

utils/utils.py

@@ -9,7 +9,7 @@ import random
 AVERAGE_SIZE = (100, 100)  # Computed with statistics
 WINDOW_SIZE_SIGN = (64, 64)
 WINDOW_SIZE_LIGHT = (43, 100)
-
+AREA_THRESHOLD = 300  # 0.05 quantile for all the areas of the training dataset, prevent to classificate too little sign/light
  
 # Dictionary for mapping class names to integers
 CLASSE_TO_INT = {