Update app.py

app.py

@@ -185,10 +185,63 @@ async def predict_single_dog(image):
         return probabilities[0], breeds[:3], relative_probs
 
 
-async def detect_multiple_dogs(image, conf_threshold=0.3, iou_threshold=0.55):
+# async def detect_multiple_dogs(image, conf_threshold=0.3, iou_threshold=0.55):
+#     results = model_yolo(image, conf=conf_threshold, iou=iou_threshold)[0]
+#     dogs = []
+#     boxes = []
+#     for box in results.boxes:
+#         if box.cls == 16:  # COCO dataset class for dog is 16
+#             xyxy = box.xyxy[0].tolist()
+#             confidence = box.conf.item()
+#             boxes.append((xyxy, confidence))
+
+#     if not boxes:
+#         dogs.append((image, 1.0, [0, 0, image.width, image.height]))
+#     else:
+#         nms_boxes = non_max_suppression(boxes, iou_threshold)
+
+#         for box, confidence in nms_boxes:
+#             x1, y1, x2, y2 = box
+#             w, h = x2 - x1, y2 - y1
+#             x1 = max(0, x1 - w * 0.05)
+#             y1 = max(0, y1 - h * 0.05)
+#             x2 = min(image.width, x2 + w * 0.05)
+#             y2 = min(image.height, y2 + h * 0.05)
+#             cropped_image = image.crop((x1, y1, x2, y2))
+#             dogs.append((cropped_image, confidence, [x1, y1, x2, y2]))
+
+#     return dogs
+
+# def non_max_suppression(boxes, iou_threshold):
+#     keep = []
+#     boxes = sorted(boxes, key=lambda x: x[1], reverse=True)
+#     while boxes:
+#         current = boxes.pop(0)
+#         keep.append(current)
+#         boxes = [box for box in boxes if calculate_iou(current[0], box[0]) < iou_threshold]
+#     return keep
+
+
+# def calculate_iou(box1, box2):
+#     x1 = max(box1[0], box2[0])
+#     y1 = max(box1[1], box2[1])
+#     x2 = min(box1[2], box2[2])
+#     y2 = min(box1[3], box2[3])
+
+#     intersection = max(0, x2 - x1) * max(0, y2 - y1)
+#     area1 = (box1[2] - box1[0]) * (box1[3] - box1[1])
+#     area2 = (box2[2] - box2[0]) * (box2[3] - box2[1])
+
+#     iou = intersection / float(area1 + area2 - intersection)
+#     return iou
+
+
+async def detect_multiple_dogs(image, conf_threshold=0.35, iou_threshold=0.55, sigma=0.5):
     results = model_yolo(image, conf=conf_threshold, iou=iou_threshold)[0]
     dogs = []
     boxes = []
+    
+    # 收集所有狗的檢測結果
     for box in results.boxes:
         if box.cls == 16:  # COCO dataset class for dog is 16
             xyxy = box.xyxy[0].tolist()
@@ -198,31 +251,69 @@ async def detect_multiple_dogs(image, conf_threshold=0.3, iou_threshold=0.55):
     if not boxes:
         dogs.append((image, 1.0, [0, 0, image.width, image.height]))
     else:
-        nms_boxes = non_max_suppression(boxes, iou_threshold)
+        # 使用SoftNMS替代原有的NMS
+        nms_boxes = soft_nms(boxes, iou_threshold, sigma)
 
+        # 處理保留的框
         for box, confidence in nms_boxes:
             x1, y1, x2, y2 = box
+            # 擴大框的範圍以包含更多上下文
             w, h = x2 - x1, y2 - y1
-            x1 = max(0, x1 - w * 0.05)
-            y1 = max(0, y1 - h * 0.05)
-            x2 = min(image.width, x2 + w * 0.05)
-            y2 = min(image.height, y2 + h * 0.05)
+            x1 = max(0, x1 - w * 0.1)  # 增加到10%的margin
+            y1 = max(0, y1 - h * 0.1)
+            x2 = min(image.width, x2 + w * 0.1)
+            y2 = min(image.height, y2 + h * 0.1)
             cropped_image = image.crop((x1, y1, x2, y2))
             dogs.append((cropped_image, confidence, [x1, y1, x2, y2]))
 
     return dogs
 
-def non_max_suppression(boxes, iou_threshold):
-    keep = []
-    boxes = sorted(boxes, key=lambda x: x[1], reverse=True)
-    while boxes:
-        current = boxes.pop(0)
-        keep.append(current)
-        boxes = [box for box in boxes if calculate_iou(current[0], box[0]) < iou_threshold]
-    return keep
-
+def soft_nms(boxes, iou_threshold=0.55, sigma=0.5, score_threshold=0.25):
+    """
+    SoftNMS with Gaussian decay
+    """
+    if not boxes:
+        return []
+    
+    # 轉換格式以便處理
+    box_coords = np.array([box[0] for box in boxes])
+    scores = np.array([box[1] for box in boxes])
+    
+    # 按照confidence排序
+    indices = np.argsort(scores)[::-1]
+    box_coords = box_coords[indices]
+    scores = scores[indices]
+    
+    keep_boxes = []
+    keep_scores = []
+    
+    while len(scores) > 0:
+        # 保留最高分數的框
+        keep_boxes.append(box_coords[0].tolist())
+        keep_scores.append(scores[0])
+        
+        if len(scores) == 1:
+            break
+        
+        # 計算當前最高分框與其他所有框的IoU
+        ious = np.array([calculate_iou(box_coords[0], box) for box in box_coords[1:]])
+        
+        # 使用高斯衰減更新分數
+        scores[1:] = scores[1:] * np.exp(-(ious * ious) / sigma)
+        
+        # 移除最高分的框並過濾低於閾值的框
+        box_coords = box_coords[1:]
+        scores = scores[1:]
+        mask = scores > score_threshold
+        box_coords = box_coords[mask]
+        scores = scores[mask]
+    
+    return list(zip(keep_boxes, keep_scores))
 
 def calculate_iou(box1, box2):
+    """
+    IoU 計算
+    """
     x1 = max(box1[0], box2[0])
     y1 = max(box1[1], box2[1])
     x2 = min(box1[2], box2[2])