README.md

---
pretty_name: "TranSpace — playroom_drjohnson"
license: other
license_name: non-commercial-research
license_link: https://github.com/codeshim/transpace/blob/main/LICENSE
tags:
  - 3d-reconstruction
  - novel-view-synthesis
  - scene-synthesis
  - indoor-scenes
size_categories:
  - 10B<n<100B
---

# TranSpace — `playroom_drjohnson` data pack

Data for the released demo of:

> **TranSpace: Progressive Anchoring for Metric-Consistent Scene Synthesis**
> Hyeshim Kim, Taehei Kim\*, Jihun Shin\*, Hyeonjin Kim, Sung-Hee Lee — *\*equal contribution*
> KAIST · ACM Multimedia 2026 (MM '26)
> [doi:10.1145/3767308.3836316](https://doi.org/10.1145/3767308.3836316)

Code: https://github.com/codeshim/transpace

TranSpace synthesizes the transition geometry between two separately captured
indoor scenes. This pack contains everything the pipeline needs to reproduce that
for the Deep Blending `playroom` and `drjohnson` rooms: the two reconstructed
rooms, the inpainted keyframes, and every intermediate video. The demo replays
these, so no generative model or API key is required.

## Download

```bash
pip install huggingface_hub

# everything (~19 GB)
hf download codeshim/transpace-playroom-drjohnson --repo-type dataset \
    --local-dir data/playroom_drjohnson
```

Place it at `data/playroom_drjohnson/` inside the code repository, then:

```bash
python main.py
```

## Contents

```
playroom_drjohnson/
├── playroom_aligned/                   room A: images/, sparse/, checkpoints/
├── drjohnson_aligned/                  room B: images/, sparse/, checkpoints/
├── BASE_model/                         floor/threshold voxel model
├── trspace_padding_model_type{0,1,2}/  space_id mask model per variant
├── trspace_source_type{0,1,2}/         inpainted keyframes + intermediate videos
└── playroom_drjohnson_alignment_info_type{0,1,2}.json
```

Each room directory is self-contained: its COLMAP reconstruction (`images/`,
`sparse/`) and its trained SVRaster voxel model (`checkpoints/`) side by side.

Three connector variants are provided for the same room pair — `type0` (wing
wall), `type1` (pony wall), `type2` (internal window). Only the RIGHT connector
varies; LEFT is an internal window in every variant. A variant's source videos,
alignment info and padding model must be used together.

`trspace_source_type*/` holds the inputs that were generated offline:

| | |
|---|---|
| `<VIEW>_generated.png` | inpainted keyframe |
| `<VIEW>_depth_raw.npy` | its aligned metric depth |
| `<START>_<END>_iterNN.mp4` | per-iteration transition video |
| `<START>_<END>_iterNN_f2m.mp4` / `_m2b.mp4` | final-closing segment videos |

## Provenance and licensing

Derived from the **Deep Blending** scenes `drjohnson` and `playroom`:

> Deep Blending for Free-Viewpoint Image-Based Rendering.
> Hedman, Philip, Price, Frahm, Drettakis, Brostow. *ACM ToG* 37(6), 2018.

The keyframes were produced with Google Vertex AI Imagen 2 (since retired) and
the intermediate videos with DeeVid AI Video Generation V2.1.

**Non-commercial use only** (research or evaluation). The voxel models are
produced by [SVRaster](https://github.com/NVlabs/svraster), whose license limits
the work and its derivatives to non-commercial use.

The paper additionally evaluates on ScanNet++ scenes, which are **not**
redistributed here.

## Citation

```bibtex
@inproceedings{kim2026transpace,
  title     = {TranSpace: Progressive Anchoring for Metric-Consistent Scene Synthesis},
  author    = {Kim, Hyeshim and Kim, Taehei and Shin, Jihun and Kim, Hyeonjin and Lee, Sung-Hee},
  booktitle = {Proceedings of the 34th ACM International Conference on Multimedia (MM '26)},
  year      = {2026},
  publisher = {ACM},
  doi       = {10.1145/3767308.3836316}
}
```