fix: skip Florence-2 detections with inverted bounding boxes

Florence-2 occasionally emits garbled location tokens that decode to
negative width/height. Filter these at parse time and guard the
annotator against any that slip through.

Author: swaynehales

uitag/annotate.py

@@ -52,6 +52,8 @@ def render_som(
 
         x1, y1 = det.x, det.y
         x2, y2 = det.x + det.width, det.y + det.height
+        if x1 > x2 or y1 > y2:
+            continue
         draw.rectangle([x1, y1, x2, y2], outline=color, width=2)
 
         # Position marker fully outside the bounding box (above-left of corner)

uitag/florence.py

@@ -79,13 +79,18 @@ def parse_location_tokens(
         x2 = int(int(match.group(4)) * image_width / 999)
         y2 = int(int(match.group(5)) * image_height / 999)
 
+        w = x2 - x1
+        h = y2 - y1
+        if w <= 0 or h <= 0:
+            continue
+
         results.append(
             {
                 "label": label,
                 "x": x1,
                 "y": y1,
-                "width": x2 - x1,
-                "height": y2 - y1,
+                "width": w,
+                "height": h,
             }
         )