app.py

import os
import gc
import cv2
import gradio as gr
import numpy as np
import matplotlib.cm as cm
import matplotlib  # New import for the updated colormap API
import subprocess
import sys
import spaces

from utils.dc_utils import read_video_frames, save_video

title = "#RGBD sbs output"
description = """**Video Depth Anything** + RGBD sbs output for viewing with Looking Glass Factory displays.
Please refer to our [paper](https://arxiv.org/abs/2501.12375), [project page](https://videodepthanything.github.io/), and [github](https://github.com/DepthAnything/Video-Depth-Anything) for more details."""

@spaces.GPU(enable_queue=True)

def stitch_rgbd_videos(
    processed_video: str,
    depth_vis_video: str,
    max_len: int = -1,
    target_fps: int = -1,
    max_res: int = 1280,
    stitch: bool = True,
    grayscale: bool = True,
    convert_from_color: bool = True,
    blur: float = 0.3,
    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)
    
    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)
        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.
        d_min, d_max = depths.min(), depths.max()
        stitched_frames = []
        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 the range [0, 255].
            depth_norm = ((depth_frame - d_min) / (d_max - d_min) * 255).astype(np.uint8)
            # Generate depth visualization:
            if grayscale:
                if convert_from_color:
                    # First, generate a color depth image using the inferno colormap,
                    # then convert that color image to grayscale.
                    cmap = matplotlib.colormaps.get_cmap("inferno")
                    depth_color = (cmap(depth_norm / 255.0)[..., :3] * 255).astype(np.uint8)
                    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.
                    depth_vis = np.stack([depth_norm] * 3, axis=-1)
            else:
                # Generate a color depth image using the inferno colormap.
                cmap = matplotlib.colormaps.get_cmap("inferno")
                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.
                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 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)
        # 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')
        save_video(stitched_frames, stitched_video_path, fps=target_fps)
        
        # Merge audio from the input video into the stitched video using ffmpeg.
        temp_audio_path = stitched_video_path.replace('_RGBD.mp4', '_RGBD_audio.mp4')
        cmd = [
            "ffmpeg",
            "-y",
            "-i", stitched_video_path,
            "-i", processed_video,
            "-c:v", "copy",
            "-c:a", "aac",
            "-map", "0:v:0",
            "-map", "1:a:0?",
            "-shortest",
            temp_audio_path
        ]
        subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
        os.replace(temp_audio_path, stitched_video_path)

    # Return stitched video.
    return [stitched_video_path]

def construct_demo():
    with gr.Blocks(analytics_enabled=False) as demo:
        gr.Markdown(title)
        gr.Markdown(description)
        gr.Markdown("### If you find this work useful, please help ⭐ the [Github Repo](https://github.com/DepthAnything/Video-Depth-Anything). Thanks for your attention!")
        
        with gr.Row(equal_height=True):
            with gr.Column(scale=1):
                # Video input component for file upload.
                processed_video = gr.Video(label="Input Video")
                depth_vis_video = gr.Video(label="Generated Depth Video")                    
            with gr.Column(scale=2):
                with gr.Row(equal_height=True):
                    stitched_video = gr.Video(label="Stitched RGBD Video", interactive=False, autoplay=True, loop=True, show_share_button=True, scale=5)
                    
        with gr.Row(equal_height=True):
            with gr.Column(scale=1):
                with gr.Accordion("Advanced Settings", open=False):
                    max_len = gr.Slider(label="Max process length", minimum=-1, maximum=1000, value=-1, step=1)
                    target_fps = gr.Slider(label="Target FPS", minimum=-1, maximum=30, value=-1, step=1)
                    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)
                    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 on display)", value=0.3)
                generate_btn = gr.Button("Generate")
            with gr.Column(scale=2):
                pass
        
        generate_btn.click(
            fn=stitch_rgbd_videos,
            inputs=[processed_video, depth_vis_video, max_len, target_fps, max_res, stitch_option, grayscale_option, convert_from_color_option, blur_slider],
            outputs=[stitched_video],
        )
    
    return demo

if __name__ == "__main__":
    demo = construct_demo()
    demo.queue(max_size=2).launch()