Update app.py

app.py

@@ -62,7 +62,8 @@ def compute_iou(boxA, boxB):
         return 0
     return interArea / float(boxAArea + boxBArea - interArea)
 
-def custom_nms(preds, iou_threshold=0.7):
+# Lowering the NMS threshold to 0.5 (from 0.7) to allow more distinct boxes when they are adjacent.
+def custom_nms(preds, iou_threshold=0.5):
     preds = sorted(preds, key=lambda x: x["confidence"], reverse=True)
     filtered_preds = []
     for pred in preds:
@@ -75,13 +76,12 @@ def custom_nms(preds, iou_threshold=0.7):
             filtered_preds.append(pred)
     return filtered_preds
 
-# In this merged process_image function we use the robust box-detection from your first prompt
-# (with updated prediction parameters) and the ArUco marker detection and size conversion
-# (modified to compute a conversion factor from the marker’s bounding box).
+# The process_image function merges robust box detection (with updated prediction parameters)
+# and ArUco marker detection with refined conversion factor computation.
 def process_image(job_id, image_path, object_type, multiplier):
     try:
         jobs[job_id]['progress'] = 10
-        # Load a fresh copy of the original image
+        # Load the original image
         image = cv2.imread(image_path)
         if image is None:
             jobs[job_id]['progress'] = 100
@@ -90,7 +90,7 @@ def process_image(job_id, image_path, object_type, multiplier):
 
         jobs[job_id]['progress'] = 20
         img_height, img_width = image.shape[:2]
-        # Compute dynamic thickness based on image size and multiplier.
+        # Set dynamic thickness based on image size and multiplier.
         thickness = max(2, int(min(img_width, img_height) / 300)) * multiplier
         detection_info = []
 
@@ -101,19 +101,19 @@ def process_image(job_id, image_path, object_type, multiplier):
                 return
 
             # --- BOX DETECTION ---
-            # Upscale the image if it is small to help detect small boxes.
+            # Upscale if image dimensions are small.
             scale_factor = 1
             if img_width < 1000 or img_height < 1000:
                 scale_factor = 2
 
-            # Use improved prediction parameters (confidence=50, overlap=30)
+            # Use improved parameters: confidence=50 and overlap=20.
             if scale_factor > 1:
                 upscaled_image = cv2.resize(image, None, fx=scale_factor, fy=scale_factor, interpolation=cv2.INTER_LINEAR)
                 temp_path = "upscaled.jpg"
                 cv2.imwrite(temp_path, upscaled_image)
-                results = box_model.predict(temp_path, confidence=50, overlap=30).json()
+                results = box_model.predict(temp_path, confidence=50, overlap=20).json()
             else:
-                results = box_model.predict(image_path, confidence=50, overlap=30).json()
+                results = box_model.predict(image_path, confidence=50, overlap=20).json()
 
             predictions = results.get("predictions", [])
             processed_preds = []
@@ -130,12 +130,12 @@ def process_image(job_id, image_path, object_type, multiplier):
                         width = prediction["width"]
                         height = prediction["height"]
 
-                    # Convert from center-based coordinates to corner-based bounding box
+                    # Convert from center coordinates to corner coordinates.
                     x1 = int(round(x - width / 2))
                     y1 = int(round(y - height / 2))
                     x2 = int(round(x + width / 2))
                     y2 = int(round(y + height / 2))
-                    # Clamp to image dimensions
+                    # Clamp to image dimensions.
                     x1 = max(0, min(x1, img_width - 1))
                     y1 = max(0, min(y1, img_height - 1))
                     x2 = max(0, min(x2, img_width - 1))
@@ -148,16 +148,14 @@ def process_image(job_id, image_path, object_type, multiplier):
                 except Exception as e:
                     continue
 
-            # Apply non-max suppression to the detected boxes
-            box_detections = custom_nms(processed_preds, iou_threshold=0.7)
+            # Apply NMS with a lower IoU threshold (0.5) to separate adjacent boxes.
+            box_detections = custom_nms(processed_preds, iou_threshold=0.5)
             jobs[job_id]['progress'] = 60
 
             # --- ARUCO MARKER DETECTION & SIZE CONVERSION ---
-            # We assume the printed marker is 5 cm x 5 cm.
-            marker_real_width_cm = 5.0  
+            marker_real_width_cm = 5.0  # The printed marker is 5 cm x 5 cm.
             try:
                 gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
-                # Use the DICT_6X6_250 dictionary (as in your current prompt)
                 aruco_dict = cv2.aruco.getPredefinedDictionary(cv2.aruco.DICT_6X6_250)
                 if hasattr(cv2.aruco, 'DetectorParameters_create'):
                     aruco_params = cv2.aruco.DetectorParameters_create()
@@ -167,7 +165,7 @@ def process_image(job_id, image_path, object_type, multiplier):
                 if ids is not None and len(corners) > 0:
                     marker_corners = corners[0].reshape((4, 2))
                     cv2.aruco.drawDetectedMarkers(image, corners, ids)
-                    # Compute the bounding box of the marker from its corners
+                    # Compute the marker's bounding box.
                     min_x = np.min(marker_corners[:, 0])
                     max_x = np.max(marker_corners[:, 0])
                     min_y = np.min(marker_corners[:, 1])
@@ -175,7 +173,7 @@ def process_image(job_id, image_path, object_type, multiplier):
                     width_pixels = max_x - min_x
                     height_pixels = max_y - min_y
                     if width_pixels > 0 and height_pixels > 0:
-                        # Use the average conversion factor from width and height
+                        # Use the average of the width and height conversion factors.
                         conversion_factor = (marker_real_width_cm / width_pixels + marker_real_width_cm / height_pixels) / 2
                     else:
                         conversion_factor = None