app.py

import spaces
import tempfile
import gradio as gr
import numpy as np
import torch
from PIL import Image
import trimesh
from huggingface_hub import hf_hub_download
from depth_anything_v2.dpt import DepthAnythingV2

css = """
#img-display-container {
    max-height: 100vh;
}
#img-display-input {
    max-height: 80vh;
}
#img-display-output {
    max-height: 80vh;
}
#download {
    height: 62px;
}
"""
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
model_configs = {
    'vits': {'encoder': 'vits', 'features': 64, 'out_channels': [48, 96, 192, 384]},
    'vitb': {'encoder': 'vitb', 'features': 128, 'out_channels': [96, 192, 384, 768]},
    'vitl': {'encoder': 'vitl', 'features': 256, 'out_channels': [256, 512, 1024, 1024]},
    'vitg': {'encoder': 'vitg', 'features': 384, 'out_channels': [1536, 1536, 1536, 1536]}
}
encoder2name = {
    'vits': 'Small',
    'vitb': 'Base',
    'vitl': 'Large',
    'vitg': 'Giant', # we are undergoing company review procedures to release our giant model checkpoint
}
encoder = 'vitl'
model_name = encoder2name[encoder]
model = DepthAnythingV2(**model_configs[encoder])
filepath = hf_hub_download(repo_id=f"depth-anything/Depth-Anything-V2-{model_name}", filename=f"depth_anything_v2_{encoder}.pth", repo_type="model")
state_dict = torch.load(filepath, map_location="cpu")
model.load_state_dict(state_dict)
model = model.to(DEVICE).eval()

title = "# Depth-Anything-V2-DepthPop"
description = """
このツールを使用すると、写真やイラストを飛び出す絵本風にすることができます。
"""
@spaces.GPU
def predict_depth(image):
    return model.infer_image(image)

def generate_point_cloud(color_img, resolution):
    depth_img = predict_depth(color_img[:, :, ::-1])
    # 画像サイズの調整
    height, width = color_img.shape[:2]
    new_height = resolution
    new_width = int(width * (new_height / height))
    
    color_img_resized = np.array(Image.fromarray(color_img).resize((new_width, new_height), Image.LANCZOS))
    depth_img_resized = np.array(Image.fromarray(depth_img).resize((new_width, new_height), Image.LANCZOS))

    # 深度の調整
    depth_min = np.min(depth_img_resized)
    depth_max = np.max(depth_img_resized)
    normalized_depth = (depth_img_resized - depth_min) / (depth_max - depth_min)
       
    # 非線形変換（必要に応じて調整）
    adjusted_depth = np.power(normalized_depth, 0.1)  # ガンマ補正
    
    # カメラの内部パラメータ（使用するカメラに基づいて調整）    
    fx, fy = 300, 300  # 焦点距離
    cx, cy = color_img_resized.shape[1] / 2, color_img_resized.shape[0] / 2  # 主点

    # メッシュグリッドの作成
    rows, cols = adjusted_depth.shape
    u, v = np.meshgrid(range(cols), range(rows))

    # 3D座標の計算（X座標を反転）
    Z = adjusted_depth
    X = -((u - cx) * Z / fx)  # X座標を反転
    Y = (v - cy) * Z / fy

    # X, Y, Z座標をスタック
    points = np.stack((X, Y, Z), axis=-1)

    # 点のリストに整形
    points = points.reshape(-1, 3)

    # 各点の色を取得
    colors = color_img_resized.reshape(-1, 3)

    # 色を0-1の範囲に正規化
    colors = colors.astype(np.float32) / 255.0

    # PointCloudオブジェクトの作成
    cloud = trimesh.PointCloud(vertices=points, colors=colors)

    # Z軸周りに180度回転を適用（時計回り）
    rotation = trimesh.transformations.rotation_matrix(np.pi, [0, 0, 1])
    cloud.apply_transform(rotation)

    # Y軸周りに180度回転を適用（上下を反転）
    flip_y = trimesh.transformations.rotation_matrix(np.pi, [0, 1, 0])
    cloud.apply_transform(flip_y)

    # GLB形式で保存
    output_path = tempfile.mktemp(suffix='.glb')
    cloud.export(output_path)

    return output_path

with gr.Blocks(css=css) as demo:
    gr.Markdown(title)
    gr.Markdown(description)
    gr.Markdown("### Depth Prediction & Point Cloud Generation")

    with gr.Row():
        input_image = gr.Image(label="Input Image", type='numpy', elem_id='img-display-input')
    with gr.Row():
        resolution_slider = gr.Slider(minimum=512, maximum=1600, value=512, step=1, label="Resolution")

    submit = gr.Button(value="Generate")
    
    output_3d = gr.Model3D(
        clear_color=[0.0, 0.0, 0.0, 0.0],
        label="3D Model",      
    )
    
    submit.click(fn=generate_point_cloud, inputs=[input_image, resolution_slider], outputs=[output_3d])

if __name__ == '__main__':
    demo.queue().launch(share=True)