Update app.py

app.py

@@ -27,9 +27,6 @@ def stitch_rgbd_videos(
     output_dir: str = './outputs',
     input_size: int = 518,
 ):
-    # 1. Read input video frames for inference (downscaled to max_res).
-    frames, target_fps = read_video_frames(processed_video, max_len, target_fps, max_res)
-    
     video_name = os.path.basename(processed_video)
     if not os.path.exists(output_dir):
         os.makedirs(output_dir)
@@ -37,7 +34,7 @@ def stitch_rgbd_videos(
     stitched_video_path = None
     if stitch:
         # For stitching: read the original video in full resolution (without downscaling).
-        full_frames, _ = read_video_frames(processed_video, max_len, target_fps, max_res=-1)
+        full_frames, target_fps = read_video_frames(processed_video, max_len, target_fps, max_res=-1)
         depths, _ = read_video_frames(depth_vis_video, max_len, target_fps, max_res=-1)
         
         # For each frame, create a visual depth image from the inferenced depths.
@@ -48,9 +45,9 @@ def stitch_rgbd_videos(
             depth_frame = depths[i]
             
             # Normalize the depth frame to the range [0, 255].
-            #depth_norm = ((depth_frame - d_min) / (d_max - d_min) * 255).astype(np.uint8)
-            depth_norm = cv2.normalize(depth_frame, None, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX)
-            depth_norm = depth_norm.astype(np.uint8)
+            depth_norm = ((depth_frame - d_min) / (d_max - d_min) * 255).astype(np.uint8)
+            #depth_norm = cv2.normalize(depth_frame, None, alpha=0, beta=255, norm_type=cv2.NORM_MINMAX)
+            #depth_norm = depth_norm.astype(np.uint8)
             
             # Generate depth visualization:
             if grayscale:
@@ -60,11 +57,13 @@ def stitch_rgbd_videos(
                     cmap = matplotlib.colormaps.get_cmap("inferno")
                     depth_color = (cmap(depth_norm / 255.0)[..., :3] * 255).astype(np.uint8)
                     
-                    if len(depth_color.shape) == 3 and depth_color.shape[2] in [3, 4]:
-                        depth_gray = cv2.cvtColor(depth_color, cv2.COLOR_RGB2GRAY)
-                    else:
-                        depth_gray = depth_color
-                    
+                    #if len(depth_color.shape) == 3 and depth_color.shape[2] in [3, 4]:
+                    #    depth_gray = cv2.cvtColor(depth_color, cv2.COLOR_RGB2GRAY)
+                    #else:
+                    #    depth_gray = depth_color                    
+                    #depth_vis = np.stack([depth_gray] * 3, axis=-1)
+
+                    depth_gray = cv2.cvtColor(depth_color, cv2.COLOR_RGB2GRAY)
                     depth_vis = np.stack([depth_gray] * 3, axis=-1)
                 else:
                     # Directly generate a grayscale image from the normalized depth values.
@@ -75,33 +74,33 @@ def stitch_rgbd_videos(
                 depth_vis = (cmap(depth_norm / 255.0)[..., :3] * 255).astype(np.uint8)
             
             # Ensure depth_vis is valid and contiguous
-            if depth_vis is None or depth_vis.size == 0:
-                raise ValueError("depth_vis is empty or not properly computed.")
-            else:
-                depth_vis = np.ascontiguousarray(depth_vis)
+            #if depth_vis is None or depth_vis.size == 0:
+            #    raise ValueError("depth_vis is empty or not properly computed.")
+            #else:
+            #    depth_vis = np.ascontiguousarray(depth_vis)
             
-            # Apply Gaussian blur if requested.
-            if blur > 0:
-                kernel_size = int(blur * 20) * 2 + 1  # Ensures an odd kernel size.
-                depth_vis = cv2.GaussianBlur(depth_vis, (kernel_size, kernel_size), 0)
+            # TODO FIX: Apply Gaussian blur if requested.
+            #if blur > 0:
+            #    kernel_size = int(blur * 20) * 2 + 1  # Ensures an odd kernel size.
+            #    depth_vis = cv2.GaussianBlur(depth_vis, (kernel_size, kernel_size), 0)
             
             # Resize the depth visualization to match the full-resolution RGB frame.
             H_full, W_full = rgb_full.shape[:2]
             depth_vis_resized = cv2.resize(depth_vis, (W_full, H_full))
             
             # Ensure both images have 3 channels.
-            if len(rgb_full.shape) == 2:
-                rgb_full = cv2.cvtColor(rgb_full, cv2.COLOR_GRAY2BGR)
-            if len(depth_vis_resized.shape) == 2:
-                depth_vis_resized = cv2.cvtColor(depth_vis_resized, cv2.COLOR_GRAY2BGR)
+            #if len(rgb_full.shape) == 2:
+            #    rgb_full = cv2.cvtColor(rgb_full, cv2.COLOR_GRAY2BGR)
+            #if len(depth_vis_resized.shape) == 2:
+            #    depth_vis_resized = cv2.cvtColor(depth_vis_resized, cv2.COLOR_GRAY2BGR)
             
             # Ensure same data type.
-            if rgb_full.dtype != depth_vis_resized.dtype:
-                depth_vis_resized = depth_vis_resized.astype(rgb_full.dtype)
+            #if rgb_full.dtype != depth_vis_resized.dtype:
+            #    depth_vis_resized = depth_vis_resized.astype(rgb_full.dtype)
             
             # Ensure images are contiguous in memory.
-            rgb_full = np.ascontiguousarray(rgb_full)
-            depth_vis_resized = np.ascontiguousarray(depth_vis_resized)
+            #rgb_full = np.ascontiguousarray(rgb_full)
+            #depth_vis_resized = np.ascontiguousarray(depth_vis_resized)
             
             # Now safely concatenate.
             stitched = cv2.hconcat([rgb_full, depth_vis_resized])