fix: bbox bugs

Signed-off-by: ivelin <[email protected]>

app.py

@@ -55,8 +55,38 @@ def process_refexp(image: Image, prompt: str):
     print(
         fr"predicted decoder sequence before token2json: {html.escape(sequence)}")
     bbox = processor.token2json(sequence)
-    bbox = bbox['target_bounding_box']
-    print(f"predicted bounding box: {bbox}")
+
+    # safeguard in case predicted sequence does not include a target_bounding_box token
+    bbox = target_bbox.get('target_bounding_box')
+    if bbox is None:
+        print(
+            f"token2bbox seq has no predicted target_bounding_box, seq:{seq}")
+        bbox = bbox = {"xmin": 0, "ymin": 0, "xmax": 0, "ymax": 0}
+        return bbox
+
+    print(f"predicted bounding box with text coordinates: {bbox}")
+    # safeguard in case text prediction is missing some bounding box coordinates
+    # or coordinates are not valid numeric values
+    try:
+        xmin = float(bbox.get("xmin", 0))
+    except ValueError:
+        xmin = 0
+    try:
+        ymin = float(bbox.get("ymin", 0))
+    except ValueError:
+        ymin = 0
+    try:
+        xmax = float(bbox.get("xmax", 1))
+    except ValueError:
+        xmax = 1
+    try:
+        ymax = float(bbox.get("ymax", 1))
+    except ValueError:
+        ymax = 1
+    # replace str with float coords
+    bbox = {"xmin": xmin, "ymin": ymin, "xmax": xmax,
+            "ymax": ymax, "decoder output sequence": sequence}
+    print(f"predicted bounding box with float coordinates: {bbox}")
 
     print(f"image object: {image}")
     print(f"image size: {image.size}")
@@ -65,23 +95,25 @@ def process_refexp(image: Image, prompt: str):
     print(f"processed prompt: {prompt}")
 
     # safeguard in case text prediction is missing some bounding box coordinates
-    xmin = math.floor(width*float(bbox.get("xmin", 0)))
-    ymin = math.floor(height*float(bbox.get("ymin", 0)))
-    xmax = math.floor(width*float(bbox.get("xmax", 1)))
-    ymax = math.floor(height*float(bbox.get("ymax", 1)))
+    xmin = math.floor(width*bbox["xmin"])
+    ymin = math.floor(height*bbox["ymin"])
+    xmax = math.floor(width*bbox["xmax"])
+    ymax = math.floor(height*bbox["ymax"])
 
     print(
         f"to image pixel values: xmin, ymin, xmax, ymax: {xmin, ymin, xmax, ymax}")
 
     shape = [(xmin, ymin), (xmax, ymax)]
 
-    # create rectangle image
+    # deaw bbox rectangle
     img1 = ImageDraw.Draw(image)
     img1.rectangle(shape, outline="green", width=5)
+    img1.rectangle(shape, outline="white", width=2)
+
     return image, bbox
 
 
-title = "Demo: Donut 🍩 for UI RefExp"
+title = "Demo: Donut 🍩 for UI RefExp (by GuardianUI)"
 description = "Gradio Demo for Donut RefExp task, an instance of `VisionEncoderDecoderModel` fine-tuned on UIBert RefExp Dataset (UI Referring Expression). To use it, simply upload your image and type a question and click 'submit', or click one of the examples to load them. Read more at the links below."
 article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2111.15664' target='_blank'>Donut: OCR-free Document Understanding Transformer</a> | <a href='https://github.com/clovaai/donut' target='_blank'>Github Repo</a></p>"
 examples = [["example_1.jpg", "select the setting icon from top right corner"],