Bounding box alignment when using overlay_instances of Visualizer class

[cris-m]

I am trying to visualize a random image from the dataset but the bounding box appear differently when using overlay_instances method and draw_dataset_dict

Here is the code:

def cv2_imshow(im):
  im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
  plt.figure(figsize=(10, 10)), plt.imshow(im), plt.axis('off');

def visualize(image, sample, colors=['r']):
   visualizer = Visualizer(image[:, :, ::-1], metadata={}, scale=1, instance_mode=ColorMode.SEGMENTATION)
   vis = visualizer.draw_dataset_dict(sample)
   cv2_imshow(vis.get_image()[:, :, ::-1])

def visualize_image_boxes(image, boxes, colors=['r']):
  v = Visualizer(image, metadata={}, scale=1, instance_mode=ColorMode.SEGMENTATION)
  v = v.overlay_instances(
      boxes = boxes,
      assigned_colors = colors * len(boxes))

  im_data = v.get_image()
  cv2_imshow(im_data[:, :, ::-1])

sample = random.choice(train_dataset_dicts)
img = cv2.imread(sample["file_name"], cv2.COLOR_RGB2BGR)
annotations = sample.get("annotations", [])

for annotation in annotations:
    bbox = annotation.get("bbox")
    segmentation = annotation.get("segmentation")

    visualize_image_boxes(img, [bbox])

visualize(img, sample)

• Using overlay_instances
  Bounding box is not well align

• Using visualize
  Bounding box is well align

The problem is from the dataset (annotation was not correct) or the code? Any help is welcome