Tutorial: Capture 4D Splats

Capturing a static room is solved. Capturing a living, breathing moment—a fountain in motion, a person moving, or a gust of wind through trees—requires 4D Gaussian Splatting (4DGS). This tutorial covers the transition from static 3D captures to dynamic volumetric time-slices.

Why Photogrammetry Fails (and 4DGS Wins)

Traditional photogrammetry and standard 3D Gaussian Splatting assume a static world. When a subject moves, the "re-projection error" spikes because the same feature point appears at different spatial coordinates across frames. This leads to the infamous "ghosting" or "blur" artifacts.

4DGS solves this by adding a temporal dimension to each Gaussian. Instead of a static position $\mu$, each splat follows a trajectory defined by a deformation field: $\mu(t) = \mu_{base} + \Delta\mu(t)$ This allows the model to track motion vectors rather than treating movement as noise, effectively "unrolling" the motion over the timeline.

Optimized Capture Workflow

1. Hardware Requirements

• Primary: iPhone 15 Pro/16 Pro (for 60FPS 4K ProRes) or a Global Shutter industrial camera.
• Stability: A DJI RS4 or equivalent gimbal. Handheld jitter is the enemy of temporal consistency.

2. Camera Trajectory Strategy

• The Orbit: Maintain a consistent radius. For 4D, you need more viewpoints per second of movement than static captures to resolve the deformation field.
• The Temporal Window: Limit your capture to 10-15 seconds. Longer captures exponentially increase the optimization complexity and VRAM requirements of the deformation field.

3. Subject Dynamics

4DGS works best with "locally rigid" motion.

• Good: A person walking, a dancer, a spinning object.
• Challenging: Smoke, flowing water, or thin transparent surfaces (which require extremely high Gaussian density and specialized loss functions).

Processing the Data

Once you have your .mov or .mp4 file, you must extract frames and camera poses (COLMAP or ARKit metadata).

# Example command for frame extraction with focus on temporal consistency
ffmpeg -i capture.mov -vf "select='not(mod(n,2))',setpts=N/FRAME_RATE/TB" -q:v 2 frames/out_%04d.png

Professional Visual Standards

Your final output should not look like "glitch art." Aim for:

• Architectural Clarity: Clean backgrounds with sharp edges.
• Data-Driven Overlays: Use depth-map visualizations to verify that your 4D deformation field is correctly mapping the subject's movement without warping the background.

Ready to render? Check out our comparison on NeRF vs. Gaussian Splatting to understand why GS is the production standard.