NammaSpace: Round 1 prototype (Techfest, IIT Bombay)

Indoor photographs → 3D reconstruction → explorable web digital twin (orbit, zoom, first-person WASD walking with collision, minimap).

Demo video (4:39, 1080p, narrated): download from Releases: NammaSpace_Round1_Demo.mp4 (full quality, 221 MB) or NammaSpace_Round1_Demo_compact.mp4 (56 MB). Videos are kept out of the git history because GitHub limits files to 100 MB.

Web viewer online: https://bhumika25157.github.io/Namma_space/ (served from the gh-pages branch, a copy of site/)

Open it on your own computer: see HOW_TO_OPEN.md.

Dataset (public, not our own capture)

Deep Blending, "Dr Johnson" scene: 263 photographs of a historic house interior, with COLMAP camera poses.

• Paper: P. Hedman et al., Deep Blending for Free-Viewpoint Image-Based Rendering, SIGGRAPH Asia 2018.
• Download used: tandt_db.zip from the 3D Gaussian Splatting release (Kerbl et al., 2023): https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/input/tandt_db.zip (folder db/drjohnson)

What's in this package

site/ web viewer + reconstructed model · pipeline/ reconstruction scripts · video/ scripts that made the demo video (narration audio and slide assets included) · docs/ · capture/ the earlier smartphone test clip. The demo videos are attached to the GitHub Release.

Left out to keep the size down (all re-downloadable or regenerable): tool binaries (tools/, 1.1 GB), the dataset (datasets/, 1 GB), intermediate reconstruction files (work/, 3.3 GB), npm packages.

Setup on a new machine

• Viewer only: just Node.js 18+ (no npm install needed). Any static web server also works.
• Rebuild / re-record (Windows + Git Bash):
  1. npm install in pipeline/ and in video/.
  2. Unpack into tools/: COLMAP 4.2 Windows no-CUDA build → tools/colmap/ (https://github.com/colmap/colmap/releases), OpenMVS 2.4 Windows release → tools/openmvs/vc17/x64/Release/ (https://github.com/cdcseacave/openMVS/releases), FFmpeg win64 GPL build → tools/ffmpeg/ffmpeg-master-latest-win64-gpl/ (https://github.com/BtbN/FFmpeg-Builds/releases).
  3. Unzip the dataset (link below) so that datasets/deepblending/db/drjohnson/{images,sparse} exists.
  4. Re-recording also needs Google Chrome (set CHROME=<path> if it isn't in the default location) and internet access for the TTS voice.

Run the viewer locally

node video/serve.mjs site 8765

Open http://localhost:8765/?scene=drjohnson

• Explore: drag = orbit · wheel = zoom (to cursor) · right-drag = pan · double-click = focus
• Walk: W A S D = move · drag = look · Shift = run · Q/E = down/up · M = minimap · K = input overlay
• Layers: Sparse / Dense / Mesh / Textured · F = fullscreen · About = data source + run statistics

Rebuild the reconstruction (Windows, CPU only; tools in tools/)

bash pipeline/reconstruct_posed.sh datasets/deepblending/db/drjohnson drjohnson   # OpenMVS densify → mesh → texture → GLB
bash pipeline/publish_drjohnson.sh                                                 # copy into site/ and register the scene

The phone-video pipeline from the earlier capture is pipeline/reconstruct.sh <video> <job>.

Re-record the video

cd video && node tts.mjs DENSE_M=<millions> && node record.mjs all && node assemble.mjs

record.mjs drives the real viewer in headless Chrome (GPU) with a virtual 30 fps clock and real mouse/keyboard events. Narration is neural TTS (Edge voice "Andrew").