Update app.py

app.py

@@ -66,6 +66,43 @@ carTransforms = transforms.Compose([transforms.Resize(224),
                                     transforms.ToTensor(),
                                     transforms.Normalize(mean=MEAN, std=STD)])
 
+                                                                        
+def cropImage(outputs, im, boxes, car_class_true):
+    # Get the masks
+    #car_class_true = outputs["instances"].pred_classes == 2
+    #boxes = list(outputs["instances"].pred_boxes[car_class_true])
+    masks = list(np.array(outputs["instances"].pred_masks[car_class_true]))
+    max_idx = torch.tensor([(x[2] - x[0])*(x[3] - x[1]) for x in boxes]).argmax().item()
+
+    # Pick an item to mask
+    item_mask = masks[max_idx]
+    # Get the true bounding box of the mask
+    segmentation = np.where(item_mask == True) # return a list of different position in the bow, which are the actual detected object
+    x_min = int(np.min(segmentation[1])) # minimum x position
+    x_max = int(np.max(segmentation[1]))
+    y_min = int(np.min(segmentation[0]))
+    y_max = int(np.max(segmentation[0]))
+    # Create cropped image from the just portion of the image we want
+    cropped = Image.fromarray(im[y_min:y_max, x_min:x_max, :], mode = 'RGB')
+    # Create a PIL Image out of the mask
+    mask = Image.fromarray((item_mask * 255).astype('uint8')) ###### change 255 
+    # Crop the mask to match the cropped image
+    cropped_mask = mask.crop((x_min, y_min, x_max, y_max))
+    # Load in a background image and choose a paste position
+    height = y_max-y_min
+    width = x_max-x_min
+    background = Image.new(mode='RGB', size=(width, height), color=(255, 255, 255, 0))
+    # Create a new foreground image as large as the composite and paste the cropped image on top
+    new_fg_image = Image.new('RGB', background.size)
+    new_fg_image.paste(cropped)
+    # Create a new alpha mask as large as the composite and paste the cropped mask
+    new_alpha_mask = Image.new('L', background.size, color=0)
+    new_alpha_mask.paste(cropped_mask)
+    #composite the foreground and background using the alpha mask
+    composite = Image.composite(new_fg_image, background, new_alpha_mask)
+    return composite
+    
+
 def classifyCar(im):
   try:
     im = cv2.imread(im)
@@ -80,12 +117,12 @@ def classifyCar(im):
     boxes = list(outputs["instances"].pred_boxes[car_class_true])
     
     if len(boxes) != 0:
-      max_idx = torch.tensor([(x[2] - x[0])*(x[3] - x[1]) for x in boxes]).argmax().item()
-      
-      im2 = Image.fromarray(np.uint8(im)).convert('RGB').crop(boxes[max_idx].to(torch.int64).numpy())
+      #max_idx = torch.tensor([(x[2] - x[0])*(x[3] - x[1]) for x in boxes]).argmax().item()
       
-      carResize = transforms.Compose([transforms.Resize((224, 224))])
-      im2 = carResize(im2)
+      #im2 = Image.fromarray(np.uint8(im)).convert('RGB').crop(boxes[max_idx].to(torch.int64).numpy())
+      im2 = cropImage(outputs, im, boxes, car_class_true)
+      #carResize = transforms.Compose([transforms.Resize((224, 224))])
+      #im2 = carResize(im2)
       
       with torch.no_grad():
         scores = torch.nn.functional.softmax(DesignModernityModel(carTransforms(im2).unsqueeze(0))[0])
@@ -106,7 +143,7 @@ def classifyCar(im):
 #examples = [[example_img.jpg], [example_img2.jpg]]  # must be uploaded in repo
 
 # create interface for model
-interface = gr.Interface(classifyCar, inputs='image', outputs=['image','label'], cache_examples=False, title='VW Up or Fiat 500')
+interface = gr.Interface(classifyCar, inputs='image', outputs=['image','label'], cache_examples=False, title='Modernity car classification')
 interface.launch()