"We are teaching machines not just to see — but to perceive, predict, and understand the living world." — Dynamic 4D Vision Project
🚀 Project Mission:
Reconstruct dynamic 4D scenes — 3D geometry evolving over time — from 2D video inputs, using state-of-the-art methods such as NeRF and 3D Gaussian Splatting.
🌍 Long-Term Goal:
Build systems that can perceive, understand, and model the real, moving world — a core step toward real-world intelligence and embodied AI.
Phase 1: Static 3D Reconstruction
- Set up PyTorch + CUDA environment
- Reproduce baseline NeRF models on standard datasets (Lego, Fern)
- Train 3D Gaussian Splatting model on toy scenes
- Capture custom real-world static scenes and reconstruct them
Phase 2: Toward Dynamic (4D) Reconstruction
- Understand and replicate dynamic scene splatting techniques
- Record own moving-object videos
- Extend models to handle time dimension
- Render dynamic, real-time moving scenes
Phase 3: Scaling Up
- Multi-view dynamic scene capture
- Real-time dynamic reconstruction optimization
- Explore interaction modeling (5D)
- Core ML: PyTorch, CUDA
- Rendering: Blender, ffmpeg
- Datasets: LLFF, Tanks & Temples, Custom captures
- Core Methods: NeRF, Gaussian Splatting, Dynamic Scene Flow
dynamic-4d-vision-reconstruction/
│
├── README.md # High-level description of project
├── notebooks/ # Jupyter notebooks for experiments
├── data/ # Pointers or scripts to download datasets
├── experiments/ # Different models, settings you try
├── models/ # Core models (NeRF, GS, etc.)
├── outputs/ # Renders, screenshots, gifs
└── resources/ # Papers, tutorials, links
I'm Yash Darji — a builder focused on advancing Spatial Intelligence and Dynamic World Modeling for the next generation of AI and robotics systems.
🔗 Twitter/X Profile
🔗 [Pearl Inc. (Coming Soon)]
- NeRF: Representing Scenes as Neural Radiance Fields
- 3D Gaussian Splatting for Real-Time Radiance Field Rendering
- Open-source pioneers whose work enables this project.