| from __future__ import annotations |
|
|
| import json |
| import os |
| import tempfile |
| from pathlib import Path |
|
|
| import gradio as gr |
| import numpy as np |
| from PIL import Image |
|
|
| from stereo_core import luminance_depth, process_video |
|
|
| VERSION = "0.1.0" |
| _depth_model = None |
|
|
|
|
| def depth_anything_fn(image: Image.Image): |
| global _depth_model |
| try: |
| import torch |
| from transformers import pipeline |
| if _depth_model is None: |
| device = 0 if torch.cuda.is_available() else -1 |
| _depth_model = pipeline("depth-estimation", model="depth-anything/Depth-Anything-V2-Small-hf", device=device) |
| result = _depth_model(image) |
| depth = np.asarray(result["depth"], dtype=np.float32) |
| depth = (depth - depth.min()) / max(float(depth.max() - depth.min()), 1e-6) |
| return 1.0 - depth |
| except Exception as exc: |
| raise RuntimeError(f"Depth Anything V2 unavailable: {exc}") from exc |
|
|
|
|
| def run(video, backend, strength, layout, max_frames): |
| if not video: |
| raise gr.Error("Upload a video first.") |
| work = Path(tempfile.mkdtemp(prefix="simam_mono2stereo_")) |
| output = work / "simam_mono2stereo.mp4" |
| if backend.startswith("Depth Anything"): |
| depth_fn = depth_anything_fn; learned = True |
| else: |
| depth_fn = luminance_depth; learned = False |
| try: |
| preview, depth, diagnostics = process_video(video, str(output), depth_fn, float(strength), layout, int(max_frames)) |
| except Exception as exc: |
| raise gr.Error(str(exc)) |
| depth_img = Image.fromarray(np.uint8(np.clip(depth, 0, 1) * 255), mode="L") |
| diagnostics.update({"project": "SimamMono2Stereo", "version": VERSION, "backend": backend, "learned_model_used": learned, "claim_boundary": "pseudo-stereo; unseen content is not verified"}) |
| return str(output), preview, depth_img, json.dumps(diagnostics, indent=2) |
|
|
|
|
| with gr.Blocks(title="SimamMono2Stereo") as demo: |
| gr.Markdown("# SimamMono2Stereo\n### Turn ordinary monocular video into an immersive stereo experiment\n\nThis is a transparent DIBR prototype. It creates a left/right pair from estimated depth and reports occlusion holes; it cannot recover unseen content like a real stereo camera.") |
| with gr.Row(): |
| with gr.Column(): |
| video = gr.Video(label="Monocular input video", sources=["upload", "webcam"], format="mp4") |
| backend = gr.Radio(["Fast luminance depth (CPU baseline)", "Depth Anything V2 Small (GPU/model download)"], value="Fast luminance depth (CPU baseline)", label="Depth backend") |
| strength = gr.Slider(0.005, 0.08, value=0.025, step=0.005, label="Stereo strength", info="Higher values increase separation and occlusion risk.") |
| layout = gr.Radio(["Side-by-side (SBS)", "Over-under (OU)"], value="Side-by-side (SBS)", label="Output layout") |
| max_frames = gr.Slider(8, 240, value=96, step=8, label="Maximum processed frames") |
| button = gr.Button("Generate immersive stereo", variant="primary") |
| gr.Examples([["examples/sample_mono.mp4"]], inputs=[video], label="Try the included synthetic fixture") |
| with gr.Column(): |
| output = gr.Video(label="Stereo video") |
| preview = gr.Image(label="First stereo frame") |
| depth = gr.Image(label="Estimated depth", image_mode="L") |
| diagnostics = gr.Code(label="Diagnostics", language="json") |
| button.click(run, [video, backend, strength, layout, max_frames], [output, preview, depth, diagnostics], api_name="generate_stereo") |
| gr.Markdown("**Research roadmap:** temporal depth → DIBR → occlusion masks → diffusion inpainting → StereoCrafter adapter → GPU H.265/WebRTC output. Pseudo-stereo quality depends on the input and is not equivalent to captured stereo.") |
|
|
|
|
| if __name__ == "__main__": |
| demo.launch(server_name=os.getenv("GRADIO_SERVER_NAME", "0.0.0.0"), server_port=int(os.getenv("GRADIO_SERVER_PORT", "7860"))) |
|
|
