Merge pull request arabian9ts#18 from arabian9ts/nms

Non Maximum Suppression

matcher.py

@@ -121,7 +121,7 @@ class label loss is evaled by loss_conf
                 matches[near_index] = Box(gt_box, gt_label)
                 pos += 1
 
-        indicies = self.extract_highest_indicies(pred_confs, pos*3)
+        indicies = self.extract_highest_indicies(pred_confs, pos*5)
 
         for i in indicies:
             if matches[i] is None and classes-1 != np.argmax(pred_confs[i]):

model/computation.py

@@ -5,6 +5,7 @@
 author: arabian9ts
 """
 
+import numpy as np
 
 def intersection(rect1, rect2):
     """
@@ -49,7 +50,7 @@ def corner2center(rect):
     center_x = (2 * rect[0] + rect[2]) * 0.5
     center_y = (2 * rect[1] + rect[3]) * 0.5
 
-    return [center_x, center_y, abs(rect[2]), abs(rect[3])]
+    return np.array([center_x, center_y, abs(rect[2]), abs(rect[3])])
 
 
 def center2corner(rect):
@@ -59,7 +60,7 @@ def center2corner(rect):
     corner_x = rect[0] - rect[2] * 0.5
     corner_y = rect[1] - rect[3] * 0.5
 
-    return [corner_x, corner_y, abs(rect[2]), abs(rect[3])]
+    return np.array([corner_x, corner_y, abs(rect[2]), abs(rect[3])])
 
 
 def convert2diagonal_points(rect):
@@ -73,7 +74,7 @@ def convert2diagonal_points(rect):
         output format is...
         [ top_left_x, top_left_y, bottom_right_x, bottom_right_y ]
     """
-    return [rect[0], rect[1], rect[0]+rect[2], rect[1]+rect[3]]
+    return np.array([rect[0], rect[1], rect[0]+rect[2], rect[1]+rect[3]])
 
 
 def convert2wh(rect):
@@ -87,4 +88,4 @@ def convert2wh(rect):
         output format is...
         [ top_left_x, top_left_y, width, height ]
     """
-    return [rect[0], rect[1], rect[2]-rect[0], rect[3]-rect[1]]
+    return np.array([rect[0], rect[1], rect[2]-rect[0], rect[3]-rect[1]])

model/default_box.py

@@ -65,7 +65,7 @@ def generate_boxes(fmap_shapes):
                 for i, ratio in enumerate(ratios):
                     s = s_k
 
-                    if 0 == i:
+                    if 1.0 == s:
                         s = np.sqrt(s_k*s_k1)
 
                     box_width = s * np.sqrt(ratio)

model/policy.py

@@ -10,15 +10,15 @@
 classes = 21
 
 # the number of boxes per feature map
-boxes = [6, 3, 6, 3, 6, 3,]
+boxes = [4, 6, 6, 6, 6, 6,]
 
 # default box ratios
 # each length should be matches boxes[index]
 box_ratios = [
+    [1.0, 1.0, 2.0, 1.0/2.0],
+    [1.0, 1.0, 2.0, 1.0/2.0, 3.0, 1.0/3.0],
+    [1.0, 1.0, 2.0, 1.0/2.0, 3.0, 1.0/3.0],
     [1.0, 1.0, 2.0, 1.0/2.0, 3.0, 1.0/3.0],
-    [1.0, 2.0, 1.0/2.0],
     [1.0, 1.0, 2.0, 1.0/2.0, 3.0, 1.0/3.0],
-    [1.0, 2.0, 1.0/2.0],
     [1.0, 1.0, 2.0, 1.0/2.0, 3.0, 1.0/3.0],
-    [1.0, 2.0, 1.0/2.0],
 ]

model/ssd.py

@@ -195,6 +195,63 @@ def jacc_filter(plabel, ploc):
                 det_labels.append(plabel)
 
         return det_locs, det_labels
+
+
+    def non_maximum_suppression(self, candidates, overlap_threshold):
+        """
+        this is nms(non maximum_suppression) which filters predicted objects.
+
+        Args:
+            predicted bounding boxes
+        Returns:
+            detected bounding boxes and its label
+        """
+
+        label = candidates[:,4]
+        boxes = candidates[label<classes-1]
+
+        if len(boxes) == 0:
+            return []
+
+        picked = []
+
+        x1 = boxes[:,0]
+        y1 = boxes[:,1]
+        x2 = boxes[:,2] + x1
+        y2 = boxes[:,3] + y1
+
+        area = (x2 - x1 + 1) * (y2 - y1 + 1)
+        idxs = np.argsort(y2)
+
+        while len(idxs) > 0:
+            last = len(idxs) - 1
+            i = idxs[last]
+            picked.append(i)
+            suppress = [last]
+
+            for pos in range(0, last):
+                j = idxs[pos]
+
+                # extract smallest and largest bouding boxes
+                xx1 = max(x1[i], x1[j])
+                yy1 = max(y1[i], y1[j])
+                xx2 = min(x2[i], x2[j])
+                yy2 = min(y2[i], y2[j])
+
+                w = max(0, xx2 - xx1 + 1)
+                h = max(0, yy2 - yy1 + 1)
+
+                # overlap of current box and those in area list
+                overlap = float(w * h) / area[j]
+
+                # suppress current box
+                if overlap > overlap_threshold:
+                    suppress.append(pos)
+
+            # delete suppressed indexes
+            idxs = np.delete(idxs, suppress)
+
+        return boxes[picked]
 
 
     def detect_objects(self, pred_confs, pred_locs):
@@ -219,9 +276,16 @@ def detect_objects(self, pred_confs, pred_locs):
         # extract top 200 by confidence
         possibilities = [np.amax(np.exp(conf)) / (np.sum(np.exp(conf)) + 1e-3) for conf in pred_confs[0]]
         indicies = np.argpartition(possibilities, -200)[-200:]
-        top200 = np.asarray(possibilities)[indicies]
-        slicer = indicies[0.7 < top200]
-
-        locations, labels = self._filter(pred_confs[0][slicer], pred_locs[0][slicer])
+        # top200 = np.asarray(possibilities)[indicies]
+        # slicer = indicies[0.7 < top200]
+        # locations, labels = self._filter(pred_confs[0][slicer], pred_locs[0][slicer])
+
+        locations, labels = pred_locs[0][indicies], np.argmax(pred_confs[0][indicies], axis=1)
+        labels = np.asarray(labels).reshape(len(labels), 1)
+        with_labels = np.concatenate((locations, labels), axis=1)
 
-        return locations, labels
+        # labels, locations = image.non_max_suppression(boxes, possibilities, 10)
+        filtered = self.non_maximum_suppression(with_labels, 0.5)
+        # locations, labels = pred_confs[0][indices], pred_locs[0][indices]
+
+        return filtered[:,:4], filtered[:,4]