Use dpt-large

app.py

@@ -4,7 +4,7 @@ from PIL import ImageDraw
 import torch
 
 detector = pipeline("zero-shot-object-detection", model="google/owlvit-base-patch32")
-depth_estimator = pipeline("depth-estimation", model="vinvino02/glpn-nyu")
+depth_estimator = pipeline("depth-estimation", model="Intel/dpt-large")
 
 def visualize_preds(image, predictions):
     new_image = image.copy()
@@ -23,21 +23,18 @@ def visualize_preds(image, predictions):
 def compute_depth(image, preds):
     
     output = depth_estimator(image)
-    print(output)
     prediction = torch.nn.functional.interpolate(
         output["predicted_depth"].unsqueeze(1),
         size=image.size[::-1],
         mode="bicubic",
         align_corners=False,
-    ).squeeze().numpy()
-
-    print(prediction.shape)
-
+    ).squeeze().cpu().numpy()
+    
     output = []
 
     for pred in preds:
-        x = (pred["box"]["xmax"] - pred["box"]["xmin"]) // 2
-        y = (pred["box"]["ymax"] - pred["box"]["ymin"]) // 2
+        x = (pred["box"]["xmax"] + pred["box"]["xmin"]) // 2
+        y = (pred["box"]["ymax"] + pred["box"]["ymin"]) // 2
         output.append({
             "class": pred["label"],
             "distance": float(prediction[y][x])