uncertain

app.py

@@ -122,7 +122,7 @@ def gpu_run_inference(predictor_arg, image, points, boxes):
     return run_inference(predictor_arg, image, points, boxes)
 
 @spaces.GPU
-def gpu_run_tracker(tracker_model_arg, tracker_viser_arg, temp_dir, video_name, grid_size, vo_points, fps):
+def gpu_run_tracker(tracker_model_arg, tracker_viser_arg, temp_dir, video_name, grid_size, vo_points, fps, mode="offline"):
     """GPU-accelerated tracking"""
     import torchvision.transforms as T
     import decord
@@ -211,7 +211,7 @@ def gpu_run_tracker(tracker_model_arg, tracker_viser_arg, temp_dir, video_name,
                             intrs=intrs, extrs=extrs, 
                             queries=query_xyt,
                             fps=1, full_point=False, iters_track=4,
-                            query_no_BA=True, fixed_cam=False, stage=1,
+                            query_no_BA=True, fixed_cam=False, stage=1, unc_metric=unc_metric,
                             support_frame=len(video_tensor)-1, replace_ratio=0.2) 
         
         # Resize results to avoid large I/O
@@ -530,7 +530,7 @@ def reset_points(original_img: str, sel_pix):
         print(f"❌ Error in reset_points: {e}")
         return None, []
 
-def launch_viz(grid_size, vo_points, fps, original_image_state):
+def launch_viz(grid_size, vo_points, fps, original_image_state, mode="offline"):
     """Launch visualization with user-specific temp directory"""
     if original_image_state is None:
         return None, None, None
@@ -560,7 +560,7 @@ def launch_viz(grid_size, vo_points, fps, original_image_state):
         out_dir = os.path.join(temp_dir, "results")
         os.makedirs(out_dir, exist_ok=True)
         
-        gpu_run_tracker(None, None, temp_dir, video_name, grid_size, vo_points, fps)
+        gpu_run_tracker(None, None, temp_dir, video_name, grid_size, vo_points, fps, mode=mode)
         
         # Process results
         npz_path = os.path.join(out_dir, "result.npz")

models/SpaTrackV2/models/predictor.py

@@ -22,8 +22,8 @@ class Predictor(torch.nn.Module):
         super().__init__()
         self.args = args
         self.spatrack = SpaTrack2(loggers=[None, None, None], **args)
-        self.S_wind = 200
-        self.overlap = 8
+        self.S_wind = args.Track_cfg.s_wind
+        self.overlap = args.Track_cfg.overlap
 
     def to(self, device: Union[str, torch.device]):
         self.spatrack.to(device)
@@ -138,7 +138,7 @@ class Predictor(torch.nn.Module):
             if extrs is not None:
                 extrs = torch.cat([extrs, extrs[-1:].repeat(T-extrs.shape[0], 1, 1)], dim=0)
             if unc_metric is not None:
-                unc_metric = torch.cat([unc_metric, unc_metric[-1:].repeat(T-unc_metric.shape[0], 1)], dim=0)
+                unc_metric = torch.cat([unc_metric, unc_metric[-1:].repeat(T-unc_metric.shape[0], 1, 1)], dim=0)
         with torch.no_grad():
             ret = self.spatrack.forward_stream(video, queries, T_org=T_,
                                                 depth=depth, intrs=intrs, unc_metric_in=unc_metric, extrs=extrs, queries_3d=queries_3d,