Update app.py

app.py

@@ -373,18 +373,33 @@ def run_pipeline(files, objects_text, depth_model_id, erosion_px,
           f"conf is None: {prediction.conf is None}; "
           f"intrinsics is None: {prediction.intrinsics is None}")
 
-    depth = prediction.depth[0]
+    def _to_numpy(t):
+        return t.cpu().numpy() if hasattr(t, "cpu") else np.asarray(t)
+
+    depth = np.squeeze(_to_numpy(prediction.depth[0])).astype(np.float32)
 
     # conf is a robustness aid, not a hard requirement -- degrade gracefully
     # to uniform confidence (median+MAD outlier rejection still applies)
     # rather than crashing if this build of DA3 doesn't return it.
     if prediction.conf is not None:
-        conf = prediction.conf[0]
+        conf = np.squeeze(_to_numpy(prediction.conf[0])).astype(np.float32)
     else:
         gr.Warning("Depth confidence map was unavailable from this DA3 build — "
                    "falling back to uniform confidence (outlier rejection still applies).")
         conf = np.ones_like(depth)
 
+    # DA3 internally resizes images before its forward pass (see the
+    # "Processed Images" log line), so depth/conf come back at that internal
+    # resolution -- not the original image's. Masks (from SAM3, via
+    # segment_batch's target_sizes=original_sizes) and intrinsics are both in
+    # original-image pixel space, so upsample depth/conf to match before
+    # indexing with the masks below (otherwise mask/depth shapes mismatch and
+    # boolean indexing raises IndexError).
+    target_w, target_h = primary_image.size
+    if depth.shape[:2] != (target_h, target_w):
+        depth = cv2.resize(depth, (target_w, target_h), interpolation=cv2.INTER_LINEAR)
+        conf = cv2.resize(conf, (target_w, target_h), interpolation=cv2.INTER_LINEAR)
+
     # Addition 2: prefer EXIF-derived intrinsics over DA3's estimated ones
     # when available and requested — a known camera beats a network guess.
     # Intrinsics are required for 3D back-projection; if DA3 didn't return