[Python] offset when visualizing masks from open-images-v7 loaded using fiftyone in python

This is how I'm getting the dataset:

train_dataset = fiftyone.zoo.load_zoo_dataset(
"open-images-v7",
split="train",
label_types=["segmentations"],
classes=["target class"],
)


This is how I'm processing and visualizing the mask for a random sample:

sample = train_dataset.take(1).first()

for detection in sample.ground_truth.detections:
if detection.label == 'target class':
bbox = detection.bounding_box
mask = detection.mask.astype(np.float32)
break

img = tf.io.read_file(sample["filepath"])
img = tf.image.decode_jpeg(img, channels=3)

x_center, y_center, w, h = bbox
image_height = img.shape[0]
image_width = img.shape[1]
# Convert to absolute coordinates
x_center_px = x_center * image_width
y_center_px = y_center * image_height
w_px = w * image_width
h_px = h * image_height
# Convert [x_center, y_center, w, h] → [x_min, y_min, x_max, y_max]
x_min = x_center_px - (w_px / 2)
y_min = y_center_px - (h_px / 2)
x_max = x_center_px + (w_px / 2)
y_max = y_center_px + (h_px / 2)

full_mask = np.zeros((image_height, image_width), dtype=np.uint8)
resized_mask = cv2.resize(mask, (int(w_px), int(h_px)), interpolation=cv2.INTER_NEAREST)
x_min, y_min = int(x_min), int(y_min)
full_mask[y_min:y_min+resized_mask.shape[0], x_min:x_min+resized_mask.shape[1]] = resized_mask

plt.imshow(img)
plt.imshow(full_mask, alpha=0.5, cmap='jet')
plt.show()


Problem:

When running above code, either it runs successfully, but in the output image, the mask always has offset and is located in upper left of the actual place, while preserving the correct size and shape.

Or, an error rises:

---> 29 full_mask[y_min:y_min+resized_mask.shape[0], x_min:x_min+resized_mask.shape[1]] = resized_mask
30
31 plt.imshow(img)

ValueError: could not broadcast input array from shape (107,512) into shape (107,0)

Continue reading...