Interactive Path Planning and Rendering for Gaussian Splatting

English | 简体中文

This tool loads a Gaussian Splatting model, lets you interactively pick a slicing plane and place control points on it, fits a smooth path, samples camera poses, filters them by visibility/overlap, and renders RGB and depth images with progress feedback.

---

Features

• Load .npz / .ply models (3DGS fields supported: f_dc_*, scale_*, rot_*, opacity)
• Interactively set a horizontal slicing plane z = const
• Place path points in order on the plane (hold Shift + Left Click)
• Fit a B‑spline and sample by spacing ds
• Filter poses by min visible points and Jaccard overlap
• Export planned_poses_*.json (intrinsics, T_wc, stats)
• Render RGB and depth (16‑bit PNG in millimeters) with progress and final summary

---

Development and Deployment

Dev mode (VS Code Dev Containers)

Prereqs: Docker, NVIDIA driver + NVIDIA Container Toolkit (GPU), VS Code + Dev Containers.

Inside the container:

cd /workspace/src/manual_splat
bash run.bash
# or
# python3 manual_plane.py --gaussians /workspace/src/data/point_cloud_gs.ply --imgw 1920 --imgh 1080

Deploy mode (run image)

Recommended script (builds if missing):

bash docker/run_all.bash \
  --gaussians $(pwd)/src/data/point_cloud_gs.ply \
  --outdir $(pwd)/output

Script: docker/run_all.bash (excerpt)

Inputs:

• --gaussians PATH: model (.ply/.npz), mounted as /mnt/gaussians.ply
• --outdir DIR: base output; script uses rgb/, depth/, poses/ subfolders

Camera: --fov, --imgw, --imgh, --near, --far

Sampling/filters: --ds, --min-visible, --overlap-ratio

Container/limits: --cpus, --mem, --omp, --safe, --runtime-nvidia, --dev

---

Launch the interactive planner

cd manual_splat
bash run.bash

bash docker/run_all.bash 

UI and interactions

• View: Left drag rotate, Right drag pan, Wheel zoom

• Plane: set Slice plane z and click Update / Show plane

  

• Draw: hold Shift + Left Click on the green plane to add points; after 2 points an orange polyline appears (drawn slightly above the plane for visibility)

• Sample: Fit spline & sample

  

• Export & render: Generate poses & export JSON & render images

  

  

  

Rendering shows real‑time progress (RGB i/N, Depth j/M). When done, a dialog summarizes counts and output locations.

---

Acknowledgements

• Open3D GUI framework
• gsplat renderer
• Gaussian Splatting paper & implementations