EVO-X2でDepth Anything 3やーる（Windows11、Python3.12）

はじめに

EVO-X2でDepth Anything 3を動かしてみました

開発環境

• EVO-X2
• Python 3.12

導入

下記を参考に、Pythonプログラムを作成

da3.py

import torch
import subprocess
import sys
import os
import time
import glob
import shutil
from datetime import datetime

# Pythonバージョンチェック
if sys.version_info < (3, 8):
    print(f"❌ Python 3.8以降が必要です。現在のバージョン: {sys.version}")
    sys.exit(1)

print(f"✅ Python {sys.version_info.major}.{sys.version_info.minor}.{sys.version_info.micro}")

# パッケージのインストール確認
required_packages = {
    'depth_anything_3': 'awesome-depth-anything-3',
    'addict': 'addict',
    'matplotlib': 'matplotlib',
    'PIL': 'pillow',
    'numpy': 'numpy',
    'torch': 'torch',
}

missing_packages = []
for module_name, package_name in required_packages.items():
    try:
        __import__(module_name)
    except ImportError:
        missing_packages.append(package_name)

if missing_packages:
    print(f"📦 Installing missing packages: {', '.join(missing_packages)}...")
    subprocess.check_call([sys.executable, "-m", "pip", "install"] + missing_packages)

# 再度インポート
from depth_anything_3.api import DepthAnything3
from depth_anything_3.utils.export.glb import export_to_glb
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import urllib.request

# ============================================================================
# 1. デバイス設定
# ============================================================================
device = "cuda" if torch.cuda.is_available() else "cpu"
gpu_name = torch.cuda.get_device_name(0) if device == "cuda" else "None"
vram = torch.cuda.get_device_properties(0).total_memory / 1e9 if device == "cuda" else 0

if device == "cuda":
    print(f"✅ GPU detected: {gpu_name} ({vram:.1f} GB VRAM)")
else:
    print("⚠️ No GPU detected. Using CPU (will be slower)")

# ============================================================================
# 2. モデルの読み込み
# ============================================================================
model_size = "DA3-LARGE"  # オプション: "DA3-SMALL", "DA3-BASE", "DA3-LARGE", "DA3-GIANT", "DA3NESTED-GIANT-LARGE"

print(f"\n📥 Loading {model_size}...")
print("| Model | Speed | Quality | VRAM |")
print("|-------|-------|---------|------|")
print("| SMALL | ⚡⚡⚡ | ★★☆ | 4GB |")
print("| BASE | ⚡⚡ | ★★★ | 6GB |")
print("| LARGE | ⚡ | ★★★★ | 8GB |")
print("| GIANT | 🐢 | ★★★★★ | 12GB |")
print("| NESTED | 🐢 | ★★★★★+ | 16GB |")

start = time.time()
model = DepthAnything3.from_pretrained(f"depth-anything/{model_size}")
model = model.to(device).eval()
print(f"✅ Model loaded in {time.time()-start:.1f}s")

# ============================================================================
# 3. サンプル画像で試す
# ============================================================================
print("\n🖼️ Testing with sample image...")

os.makedirs("samples", exist_ok=True)
url = "https://images.unsplash.com/photo-1506905925346-21bda4d32df4?w=1280"
sample_path = "samples/mountain.jpg"

if not os.path.exists(sample_path):
    print("📥 Downloading sample image...")
    urllib.request.urlretrieve(url, sample_path)

# 推論実行
result = model.inference([sample_path])

# 可視化
fig, axes = plt.subplots(1, 2, figsize=(14, 5))

axes[0].imshow(result.processed_images[0])
axes[0].set_title("📸 Input", fontsize=14, fontweight='bold')
axes[0].axis("off")

depth = result.depth[0]
im = axes[1].imshow(depth, cmap='Spectral_r')
axes[1].set_title(f"🌊 Depth (range: {depth.min():.1f}m - {depth.max():.1f}m)", fontsize=14, fontweight='bold')
axes[1].axis("off")
plt.colorbar(im, ax=axes[1], fraction=0.046, pad=0.04, label='Depth (m)')

plt.tight_layout()
plt.savefig("samples/depth_result.png", dpi=150, bbox_inches='tight')
print("💾 Saved visualization to samples/depth_result.png")
plt.close()

print(f"\n📊 Output shapes:")
print(f"   Depth: {result.depth.shape}")
print(f"   Confidence: {result.conf.shape}")
print(f"   Camera intrinsics: {result.intrinsics.shape}")

# ============================================================================
# 4. 独自の画像で3D再構成
# ============================================================================
print("\n" + "="*60)
print("📤 4. Use Your Own Images")
print("="*60)

# 画像ディレクトリを指定（必要に応じて変更）
upload_dir = "my_images"  # または画像があるディレクトリのパスを指定

print(f"\n💡 Tips for best results:")
print("   - Move the camera, not the objects")
print("   - 30-50% overlap between consecutive images")
print("   - Avoid motion blur")
print("   - Good lighting helps!")

if not os.path.exists(upload_dir):
    print(f"\n❌ Directory '{upload_dir}' not found.")
    print(f"Please create the directory and add your images there.")
    print(f"   mkdir {upload_dir}")
    image_files = []
else:
    image_files = sorted(glob.glob(f"{upload_dir}/*.jpg") + 
                        glob.glob(f"{upload_dir}/*.jpeg") + 
                        glob.glob(f"{upload_dir}/*.png") + 
                        glob.glob(f"{upload_dir}/*.webp"))

    print(f"\n✅ Found {len(image_files)} images in '{upload_dir}'")

    # プレビュー
    if len(image_files) > 0:
        n_preview = min(6, len(image_files))
        fig, axes = plt.subplots(1, n_preview, figsize=(3*n_preview, 3))
        if n_preview == 1:
            axes = [axes]
        for i, img_path in enumerate(image_files[:n_preview]):
            img = Image.open(img_path)
            axes[i].imshow(img)
            axes[i].set_title(f"#{i+1}", fontsize=10)
            axes[i].axis("off")
        if len(image_files) > n_preview:
            print(f"   (showing first {n_preview} of {len(image_files)})")
        plt.tight_layout()
        plt.savefig("samples/image_preview.png", dpi=150, bbox_inches='tight')
        plt.close()

# 3D再構成
if len(image_files) >= 2:
    print(f"\n🔄 Processing {len(image_files)} images...")
    start = time.time()

    result = model.inference(
        image_files,
        process_res_method="upper_bound_resize",
    )

    inference_time = time.time() - start
    print(f"✅ Inference done in {inference_time:.1f}s ({len(image_files)/inference_time:.1f} img/s)")

    # GLBにエクスポート
    output_dir = "output_3d"
    os.makedirs(output_dir, exist_ok=True)

    print("📦 Generating 3D point cloud...")
    export_to_glb(
        result,
        export_dir=output_dir,
        show_cameras=True,
        conf_thresh_percentile=20,  # Filter low-confidence points
        num_max_points=500_000,
    )

    print(f"\n✅ 3D model saved to {output_dir}/")
    if os.path.exists(output_dir):
        files = os.listdir(output_dir)
        for f in files:
            filepath = os.path.join(output_dir, f)
            size = os.path.getsize(filepath) / (1024 * 1024)  # MB
            print(f"  📄 {f}: {size:.2f} MB")

    glb_file = os.path.join(output_dir, "point_cloud.glb")
    if os.path.exists(glb_file):
        print(f"\n🎉 GLB file created!")
        print(f"📁 Location: {os.path.abspath(glb_file)}")
        print(f"\n👉 View your model: https://gltf-viewer.donmccurdy.com/")
elif len(image_files) == 1:
    print("\n⚠️ At least 2 images are needed for 3D reconstruction.")
else:
    print("\n⚠️ No images found. Skipping 3D reconstruction.")

# ============================================================================
# 5. 結果の可視化
# ============================================================================
if 'result' in locals() and len(result.depth) > 0 and len(result.depth) > 1:
    print("\n" + "="*60)
    print("📊 5. Visualize Results")
    print("="*60)

    # 深度マップの表示
    n_images = len(result.depth)
    cols = min(4, n_images)
    rows = (n_images + cols - 1) // cols

    fig, axes = plt.subplots(rows, cols, figsize=(4*cols, 4*rows))
    axes = np.array(axes).flatten() if n_images > 1 else [axes]

    for i in range(n_images):
        depth = result.depth[i]
        axes[i].imshow(depth, cmap='Spectral_r')
        axes[i].set_title(f"Frame {i+1}", fontsize=10)
        axes[i].axis("off")

    # 未使用のサブプロットを非表示
    for i in range(n_images, len(axes)):
        axes[i].axis("off")

    plt.suptitle("🌊 Depth Maps", fontsize=16, fontweight='bold')
    plt.tight_layout()
    plt.savefig("samples/all_depth_maps.png", dpi=150, bbox_inches='tight')
    print("💾 Saved depth maps to samples/all_depth_maps.png")
    plt.close()

    # カメラポーズの可視化
    if len(result.extrinsics) > 0:
        from mpl_toolkits.mplot3d import Axes3D

        positions = []
        for ext in result.extrinsics:
            R = ext[:3, :3]
            t = ext[:3, 3]
            cam_pos = -R.T @ t
            positions.append(cam_pos)

        positions = np.array(positions)

        fig = plt.figure(figsize=(10, 8))
        ax = fig.add_subplot(111, projection='3d')

        ax.scatter(positions[:, 0], positions[:, 1], positions[:, 2],
                   c=range(len(positions)), cmap='viridis', s=100, marker='o')
        ax.plot(positions[:, 0], positions[:, 1], positions[:, 2],
                'b-', alpha=0.5, linewidth=1)
        ax.scatter(*positions[0], c='green', s=200, marker='^', label='First')
        ax.scatter(*positions[-1], c='red', s=200, marker='v', label='Last')

        ax.set_xlabel('X')
        ax.set_ylabel('Y')
        ax.set_zlabel('Z')
        ax.set_title('📷 Camera Trajectory', fontsize=14, fontweight='bold')
        ax.legend()

        plt.tight_layout()
        plt.savefig("samples/camera_trajectory.png", dpi=150, bbox_inches='tight')
        print("💾 Saved camera trajectory to samples/camera_trajectory.png")
        plt.close()

        print(f"📍 {len(positions)} camera poses estimated")

print("\n" + "="*60)
print("✅ Processing complete!")
print("="*60)

ライブラリのインストール

pip install awesome-depth-anything-3 matplotlib pillow numpy torch

画像を my_images/ フォルダに配置

mkdir my_images

000.png

010.png

実行

実行してみましょう

python da3.py

✅ Python 3.12.12
[INFO ] ModelCache initialized
📦 Installing missing packages: awesome-depth-anything-3, addict...
[WARN ] Dependency `gsplat` is required for rendering 3DGS. Install via: pip install git+https://github.com/nerfstudio-project/gsplat.git@0b4dddf04cb687367602c01196913cde6a743d70
⚠️ No GPU detected. Using CPU (will be slower)

📥 Loading DA3-LARGE...
| Model | Speed | Quality | VRAM |
|-------|-------|---------|------|
| SMALL | ⚡⚡⚡ | ★★☆ | 4GB |
| BASE | ⚡⚡ | ★★★ | 6GB |
| LARGE | ⚡ | ★★★★ | 8GB |
| GIANT | 🐢 | ★★★★★ | 12GB |
| NESTED | 🐢 | ★★★★★+ | 16GB |
config.json: 1.21kB [00:00, 1.21MB/s]
[INFO ] Model cache MISS: da3-large on cpu. Loading...
[INFO ] using MLP layer as FFN
[INFO ] Model cached: da3-large on cpu
[INFO ] Using standard InputProcessor (optimized CPU pipeline)
model.safetensors: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.64G/1.64G [01:29<00:00, 18.3MB/s]
✅ Model loaded in 92.7s

🖼️ Testing with sample image...
📥 Downloading sample image...
[INFO ] Processed Images Done taking 0.07385778427124023 seconds. Shape:  torch.Size([1, 3, 336, 504])
[INFO ] Model Forward Pass Done. Time: 5.336712837219238 seconds
[INFO ] Conversion to Prediction Done. Time: 0.0010001659393310547 seconds
💾 Saved visualization to samples/depth_result.png

📊 Output shapes:
   Depth: (1, 336, 504)
   Confidence: (1, 336, 504)
   Camera intrinsics: (1, 3, 3)

サンプル画像の深度推定結果

============================================================
📤 4. Use Your Own Images
============================================================

💡 Tips for best results:
   - Move the camera, not the objects
   - 30-50% overlap between consecutive images
   - Avoid motion blur
   - Good lighting helps!

✅ Found 2 images in 'my_images'

🔄 Processing 2 images...
[INFO ] Processed Images Done taking 0.02064800262451172 seconds. Shape:  torch.Size([2, 3, 280, 504])
[INFO ] Model Forward Pass Done. Time: 8.261384725570679 seconds
[INFO ] Conversion to Prediction Done. Time: 0.0 seconds
✅ Inference done in 8.3s (0.2 img/s)
📦 Generating 3D point cloud...
[INFO ] conf_thresh_percentile: 20
[INFO ] num max points: 500000
[INFO ] Exporting to GLB with num_max_points: 500000
'cp' は、内部コマンドまたは外部コマンド、
操作可能なプログラムまたはバッチ ファイルとして認識されていません。

✅ 3D model saved to output_3d/
  📄 depth_vis: 0.00 MB
  📄 scene.glb: 2.07 MB

============================================================
📊 5. Visualize Results
============================================================
💾 Saved depth maps to samples/all_depth_maps.png
💾 Saved camera trajectory to samples/camera_trajectory.png
📍 2 camera poses estimated

============================================================
✅ Processing complete!
============================================================

my_images/に配置した2枚の画像

深度推定結果
output_3d/depth_vis

0000.jpg

0001.jpg

samples/

カメラの軌跡

2枚の画像から3D再構築したモデル
output_3d/scene.glb

glbビューワー作った pic.twitter.com/HD55CzGsNr

— がちもとさん (@sotongshi) December 9, 2025

お疲れさまでした。