Update app.py

app.py

@@ -12,19 +12,19 @@ from utils.dc_utils import read_video_frames, save_video
 from huggingface_hub import hf_hub_download
 
 # Examples for the Gradio Demo.
-# Each example now contains 8 parameters:
-# [video_path, max_len, target_fps, max_res, stitch, grayscale, blur, convert_from_color]
+# Each example now contains 8 parameters in the following order:
+# [video_path, max_len, target_fps, max_res, stitch, grayscale, convert_from_color, blur]
 examples = [
-    ['assets/example_videos/davis_rollercoaster.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/Tokyo-Walk_rgb.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/4158877-uhd_3840_2160_30fps_rgb.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/4511004-uhd_3840_2160_24fps_rgb.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/1753029-hd_1920_1080_30fps.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/davis_burnout.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/example_5473765-l.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/Istanbul-26920.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/obj_1.mp4', -1, -1, 1280, True, True, 0, True],
-    ['assets/example_videos/sheep_cut1.mp4', -1, -1, 1280, True, True, 0, True],
+    ['assets/example_videos/davis_rollercoaster.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/Tokyo-Walk_rgb.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/4158877-uhd_3840_2160_30fps_rgb.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/4511004-uhd_3840_2160_24fps_rgb.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/1753029-hd_1920_1080_30fps.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/davis_burnout.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/example_5473765-l.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/Istanbul-26920.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/obj_1.mp4', -1, -1, 1280, True, True, True, 0.5],
+    ['assets/example_videos/sheep_cut1.mp4', -1, -1, 1280, True, True, True, 0.5],
 ]
 
 # Use GPU if available; otherwise, use CPU.
@@ -63,11 +63,10 @@ def infer_video_depth(
     max_res: int = 1280,
     stitch: bool = True,
     grayscale: bool = True,
-    blur: float = 0.0,
-    *,  # The following parameters are keyword-only (not overridden by UI input)
+    convert_from_color: bool = True,
+    blur: float = 0.5,
     output_dir: str = './outputs',
     input_size: int = 518,
-    convert_from_color: bool = True,
 ):
     # 1. Read input video frames for inference (downscaled to max_res).
     frames, target_fps = read_video_frames(input_video, max_len, target_fps, max_res)
@@ -94,7 +93,7 @@ def infer_video_depth(
         for i in range(min(len(full_frames), len(depths))):
             rgb_full = full_frames[i]  # Full-resolution RGB frame.
             depth_frame = depths[i]
-            # Normalize the depth frame to [0, 255].
+            # Normalize the depth frame to the range [0, 255].
             depth_norm = ((depth_frame - d_min) / (d_max - d_min) * 255).astype(np.uint8)
             # Generate depth visualization:
             if grayscale:
@@ -113,16 +112,16 @@ def infer_video_depth(
                 depth_vis = (cmap(depth_norm / 255.0)[..., :3] * 255).astype(np.uint8)
             # Apply Gaussian blur if requested.
             if blur > 0:
-                kernel_size = int(blur * 20) * 2 + 1  # ensures an odd kernel size.
+                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))
-            # Concatenate full-resolution RGB (left) and resized depth visualization (right).
+            # Concatenate the full-resolution RGB frame (left) and the resized depth visualization (right).
             stitched = cv2.hconcat([rgb_full, depth_vis_resized])
             stitched_frames.append(stitched)
         stitched_frames = np.array(stitched_frames)
-        # Limit the video name to the first 20 characters and append '_RGBD.mp4'
+        # Use only the first 20 characters of the base name for the output filename and append '_RGBD.mp4'
         base_name = os.path.splitext(video_name)[0]
         short_name = base_name[:20]
         stitched_video_path = os.path.join(output_dir, short_name + '_RGBD.mp4')
@@ -175,15 +174,15 @@ def construct_demo():
                     max_res = gr.Slider(label="Max side resolution", minimum=480, maximum=1920, value=1280, step=1)
                     stitch_option = gr.Checkbox(label="Stitch RGB & Depth Videos", value=True)
                     grayscale_option = gr.Checkbox(label="Output Depth as Grayscale", value=True)
-                    blur_slider = gr.Slider(minimum=0, maximum=1, step=0.01, label="Depth Blur Factor", value=0)
                     convert_from_color_option = gr.Checkbox(label="Convert Grayscale from Color", value=True)
+                    blur_slider = gr.Slider(minimum=0, maximum=1, step=0.01, label="Depth Blur (can reduce edge artifacts)", value=0.5)
                 generate_btn = gr.Button("Generate")
             with gr.Column(scale=2):
                 pass
         
         gr.Examples(
             examples=examples,
-            inputs=[input_video, max_len, target_fps, max_res, stitch_option, grayscale_option, blur_slider, convert_from_color_option],
+            inputs=[input_video, max_len, target_fps, max_res, stitch_option, grayscale_option, convert_from_color_option, blur_slider],
             outputs=[processed_video, depth_vis_video, stitched_video],
             fn=infer_video_depth,
             cache_examples=True,
@@ -192,7 +191,7 @@ def construct_demo():
         
         generate_btn.click(
             fn=infer_video_depth,
-            inputs=[input_video, max_len, target_fps, max_res, stitch_option, grayscale_option, blur_slider, convert_from_color_option],
+            inputs=[input_video, max_len, target_fps, max_res, stitch_option, grayscale_option, convert_from_color_option, blur_slider],
             outputs=[processed_video, depth_vis_video, stitched_video],
         )