3D Reconstruction (Scene) + Semantic GS with AI: Complete Python Guide (DepthAnything v3)
GLTF Viewer → https://learngeodata.eu/apps/viewers/...
Get 3D Assets → https://learngeodata.eu/3d-newsletter/
Traditional photogrammetry requires high overlap, complex calibration, and heavy computation. This blueprint flips the script:
Input: Random, uncalibrated 2D images (even from a rainy day).
Process: AI-driven metric depth estimation + semantic geometry projection.
Output: Dense 3D Point Clouds & Gaussian Splatting compatible sets.
🛠️ Arsenal (Requirements)
You need a GPU. Don't try this on a CPU unless you like waiting.
torch & torchvision (CUDA 11.8+ recommended)
open3d (for visualization and point cloud handling)
cv2 (OpenCV)
depth_anything_v3 (The star of the show)
⏱️ TIMESTAMPS:
[00:00] The Rainy Car Experiment
[01:15] Why Traditional Photogrammetry Failed
[03:30] The Data: 13 Uncalibrated Images
[05:45] Setting Up the Python Environment
[07:20] Depth Anything V3 Architecture
[09:50] Loading the Model & Pre-trained Weights
[12:10] Step 1: Image Ingestion & Preprocessing
[14:40] Step 2: Running the Depth Inference
[17:15] Step 3: Inverse Projection (2D to 3D)
[20:00] Visualizing the Raw Point Cloud
[22:30] Optimizing Performance (3ms Speed)
[24:15] Filtering Noise & Outliers
[26:00] Exporting to Gaussian Splatting
[27:45] The Future of 3D AI & Conclusion
🙋 FOLLOW ME
Linkedin: / florent-poux-point-cloud
Medium: / florentpoux
#3DRendering #GaussianSplatting #BlenderPython #SyntheticData #3DAI #ComputerVision #MachineLearning
WHO AM I?
If we haven’t yet before - Hey 👋 I’m Florent, a professor-turned-entrepreneur, and I’ve somehow become one of the most-followed 3D experts. Through my videos here on this channel and my writing, I share evidence-based strategies and tools to help you be better coders and 3D innovators.