Optimize model loading to prevent reload loop

inference.py

@@ -11,7 +11,7 @@ from threading import RLock
 from models.detectors.base import ObjectDetector
 from models.model_loader import load_detector, load_detector_on_device
 from models.segmenters.model_loader import load_segmenter, load_segmenter_on_device
-from models.depth_estimators.model_loader import load_depth_estimator_on_device
+from models.depth_estimators.model_loader import load_depth_estimator, load_depth_estimator_on_device
 from utils.video import extract_frames, write_video
 
 
@@ -529,6 +529,15 @@ def run_inference(
 
     else:
         # Standard Single-Threaded Loop
+        # Pre-load models to ensure they are loaded once
+        detector_instance = load_detector(active_detector)
+        detector_instance.lock = _get_model_lock("detector", detector_instance.name)
+        
+        depth_estimator_instance = None
+        if depth_estimator_name:
+            depth_estimator_instance = load_depth_estimator(depth_estimator_name)
+            depth_estimator_instance.lock = _get_model_lock("depth", depth_estimator_instance.name)
+
         processed_frames = []
         all_detections = []
         for idx, frame in enumerate(frames):
@@ -540,14 +549,17 @@ def run_inference(
             logging.debug("Processing frame %d", idx)
             
             # Run depth estimation every 3 frames if configured
-            active_depth = depth_estimator_name if (idx % 3 == 0) else None
+            active_depth_name = depth_estimator_name if (idx % 3 == 0) else None
+            active_depth_instance = depth_estimator_instance if (idx % 3 == 0) else None
             
             processed_frame, frame_dets = infer_frame(
                 frame, 
                 queries, 
-                detector_name=active_detector,
-                depth_estimator_name=active_depth,
-                depth_scale=depth_scale
+                detector_name=None,
+                depth_estimator_name=active_depth_name,
+                depth_scale=depth_scale,
+                detector_instance=detector_instance,
+                depth_estimator_instance=active_depth_instance
             )
             processed_frames.append(processed_frame)
             all_detections.append(frame_dets)