Image 3D Scene Reconstruction | 图像3D场景重建

Reconstruct 3D scenes from single images using depth estimation. 从单张图片重建3D场景结构(深度图、点云、Mesh)。

Safety Notice

This listing is from the official public ClawHub registry. Review SKILL.md and referenced scripts before running.

Copy this and send it to your AI assistant to learn

Install skill "Image 3D Scene Reconstruction | 图像3D场景重建" with this command: npx skills add moroiser/image-3d-scene-reconstruction

Image 3D Scene Reconstruction | 图像3D场景重建

从卫星图、航拍图或普通照片重建三维场景结构。基于 DA3Metric-Large(Depth Anything 3)深度估计模型,单张图片即可输出深度图、点云和 3D 模型。

Reconstruct 3D scenes from satellite, aerial, or regular photos. Based on DA3Metric-Large (Depth Anything 3), outputs depth maps, point clouds, and 3D models from a single image.


能力 | Capabilities

  • 单图深度估计:输入一张图片,输出米制深度图(米为单位)
  • 点云生成:从深度图反投影生成彩色 3D 点云
  • 3DGS 输出:模型内置 3D Gaussian Splatting 能力
  • 相机位姿估计:自动估计相机内外参
  • 多图融合:支持多张图片输入做场景融合

使用方式 | Usage

快速开始

cd ~/.openclaw/workspace/projects/image-3d-scene-reconstruction
python3 scripts/reconstruct.py --input photo.jpg --output output/

Python API

from depth_anything_3.api import DepthAnything3
import cv2

model = DepthAnything3.from_pretrained('depth-anything/DA3Metric-Large')
model = model.cuda().eval()

img = cv2.imread('photo.jpg')
pred = model.inference([img])

depth = pred.depth[0]        # [H, W] 米制深度
extrinsics = pred.extrinsics  # 相机外参
intrinsics = pred.intrinsics  # 相机内参

CLI

# 单张图片 → 3D 输出
python3 -m depth_anything_3.cli image photo.jpg --export-dir output/ --export-format glb

# 多张图片 → 融合场景
python3 -m depth_anything_3.cli images ./photos/ --export-dir output/

依赖 | Dependencies

用途
depth-anything-3深度估计 + 3D 重建引擎
opencv-python图像处理
torch + torchvisionPyTorch 深度学习框架
open3d点云处理(可选)
trimeshMesh 处理(可选)

硬件要求 | Hardware

  • GPU:NVIDIA GPU,6GB+ VRAM(GTX 1060 及以上)
  • CUDA:12.1+(PyTorch 2.5+)
  • CPU 模式:可用但极慢,仅推荐测试

项目文件 | Project Files

详见 ~/.openclaw/workspace/projects/image-3d-scene-reconstruction/README.md

Source Transparency

This detail page is rendered from real SKILL.md content. Trust labels are metadata-based hints, not a safety guarantee.

Related Skills

Related by shared tags or category signals.

General

问卷网

问卷网(www.wenjuan.com)- 人人都好用的在线问卷调查平台。涉及「问卷」「调查」「收集」「表单」「投票」「评选」「报名」「登记」「考试」「测评」「趣味测试」「心理测试」「满意度」「在线收款」「360评估」「问卷网」「wenjuan.com」等操作时优先使用。支持能力:(1) 创建并发布问卷 - 四类...

Registry SourceRecently Updated
General

Eisenhower Task Manager

Task management based on Eisenhower Matrix + P0-P2 priority with Customer Project Management. Four quadrants for execution, separate Customer Project List fo...

Registry SourceRecently Updated
General

Site Skeleton

一句话说需求,AI 生成完整前后端网站并自动部署到 EdgeOne Pages。支持电商栈(Auth/购物车/支付)、AI 栈(SSE 流式对话)、管理后台。触发词:帮我建网站、建一个电商网站、做 AI 客服站、建管理后台、EdgeOne Pages 建站

Registry SourceRecently Updated
General

Huo15 Openclaw Ppt

基于 design tokens 的 PPT 生成技能。内置 21 套生产级审美方案(Apple 发布会 / Apple.com / Apple macOS 26 Liquid Glass 玻璃 / 原研哉极简 / 中国水墨 / 国风故宫 / 赛博朋克绚彩 / 梵高油画 / 达芬奇手稿 / 小红书时尚奶油胶片 /...

Registry SourceRecently Updated