File size: 6,702 Bytes
6238ee9
ca5c68e
 
 
 
 
 
 
 
 
6238ee9
ca5c68e
 
 
 
 
 
46f479f
 
 
ca5c68e
 
46f479f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ca5c68e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46f479f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ca5c68e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f9d89f8
 
 
 
 
 
 
 
ca5c68e
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
---
pretty_name: Fire3D Single-Image Data
size_categories: n<1K
tags:
  - 3d
  - scene-reconstruction
  - single-image-3d
  - point-cloud
  - 3d-front
  - depth
---

# Fire3D Single-Image Data

20 indoor scenes rendered from [3D-FRONT](https://tianchi.aliyun.com/specials/promotion/alibaba-3d-scene-dataset).
Each scene is one RGB view with metric depth, an empty-room depth map, ground-truth object
meshes, and per-object 3D boxes — for single-image 3D scene reconstruction and evaluation.
19 of the 20 scenes additionally ship a **full-resolution instance mask** and **complete
per-object / whole-scene meshes** recovered from the source 3D-FRONT room (see
[Instance-level GT](#instance-level-gt-from-3d-front)). ~305 MiB.

```
<index>/                            # e.g. 3025, index06 = 003025
├── rgb_<index06>.jpeg              # 1296x968 RGB
├── depth_<index06>.npy             # (968, 1296) float64, metric z-depth
├── bgdepth_<index06>.npy           # (484, 648)  float64, depth of the empty room (half res)
├── annotation_<index06>.json       # intrinsics, extrinsics, per-object boxes + labels
├── sceneobjgt_<index06>.ply        # partial GT object meshes (as shipped), OpenCV camera frame

│   # instance-level GT aligned from 3D-FRONT (19/20 scenes; absent for 3966):
├── instance_<index06>.png          # (968, 1296) uint16 instance-id mask (0 = background)
├── instance_<index06>.json         # id -> {label, obj_id, model_uuid, n_pixels, ...} + match/similarity
├── instance_overlay_<index06>.png  # RGB with the coloured mask blended in, for eyeballing
├── sceneobjfull_<index06>.ply      # all objects merged, OpenCV camera frame
└── objects_<index06>/              # per-object meshes: <instid:03d>_<label>.ply
lift_3d_pcd.py                      # reference loader: depth -> point cloud, + viewer
build_gt_from_3dfront.py            # regenerates the instance-level GT from 3D-FRONT
gt_3dfront_summary.json             # per-scene match report (room, depth-agreement stats)
```

Scenes: `3025 3084 3126 3200 3266 3277 3376 3392 3401 3431 3454 3477 3844 3847 3966 4033 4087 4091 4124 4135`

## Download

```bash
pip install -U "huggingface_hub[cli]"
hf download TianhangCheng7/Fire3DSingleImageData --repo-type dataset --local-dir single_image

# one scene
hf download TianhangCheng7/Fire3DSingleImageData --repo-type dataset \
  --local-dir single_image --include "3025/*"
```

Resumable — re-run if interrupted. In Python: `snapshot_download("TianhangCheng7/Fire3DSingleImageData", repo_type="dataset")`.

## Conventions

- `depth` is **z-depth** (along the optical axis), in metres. Windows / sky are stored as
  ~9e3 — mask with `depth < 100`.
- `bgdepth` is the same view without objects, at **half resolution**; upsample before
  comparing with `depth`.
- `camera_intrinsics` is a pinhole `K` for the **full** 968×1296 image, with integer pixel
  indices (`u` = column, `v` = row).
- `camera_extrinsics` = `[R|t]` maps world → **OpenGL** camera (x right, y up, z back) and
  is the frame of `bbox3d_camera`. `camera_pose_rot` / `camera_pose_tran` are its inverse.
- `sceneobjgt_*.ply` is in the **OpenCV** camera frame (x right, y down, z forward) — the
  frame you get by unprojecting `depth` with `K`, so it overlays the lifted cloud directly.
- World frame is 3D-FRONT: z up, floor at z = 0.
- 2D boxes (`bbox_2d`, `bbox_2d_from_3d`, `render_box`) are at **half resolution** —
  multiply by 2 for full-res pixels.

Each `obj_dict` entry has `label` / `cls_id`, `obj_id` / `model_file_name`,
`obj_tran` / `obj_rot` (local→world) / `obj_scale`, `bbox3d_world` (3, 8) corners,
`bbox3d_world_center` + `half_length`, `bbox3d_camera`, the 2D boxes, and `occ_iou`.

## Instance-level GT (from 3D-FRONT)

Because `single_image` is a subset of 3D-FRONT, each view is re-aligned to its source
3D-FRONT room to recover a **complete** instance mask and object meshes (the shipped
`sceneobjgt_*.ply` is only a partial merge). `build_gt_from_3dfront.py` does this:

1. index each furniture model UUID → the 3D-FRONT rooms containing it (from the
   `*_full.glb` scene graphs);
2. shortlist rooms whose models cover the annotation and RANSAC-fit the glb→world
   similarity transform from object-centroid correspondences;
3. disambiguate the room by ray-casting the placed objects through the annotation
   intrinsics and comparing the rendered z-depth against the metric `depth`;
4. write the outputs above into the scene folder.

Objects only — walls / floor / ceiling are excluded. The meshes are in the **OpenCV**
camera frame (same as `sceneobjgt_*` / the lifted cloud), so `sceneobjfull_*.ply` overlays
`depth` directly. `instance_*.png` is full resolution and aligned to `rgb_*`; ids `1..N`
index the `objects_<index06>/` meshes and the `instances` list in `instance_*.json`.

Coverage: **19/20 scenes** align to < 1.5 mm median depth error (> 87 % of object pixels
within 5 mm; see `gt_3dfront_summary.json`). **3966 is intentionally omitted** — its
objects are near-coplanar (six identical chairs) so the room could not be recovered
reliably; it keeps only the originally shipped files.

Regenerate (needs `pip install trimesh embreex pillow`; uses a model-UUID index built from
the local 3D-FRONT copy):

```bash
python build_gt_from_3dfront.py                 # all scenes
python build_gt_from_3dfront.py --scenes 3025   # one scene
```

## Point clouds and visualization

`lift_3d_pcd.py` sits in the dataset root and finds the scenes next to itself, so it runs
with no configuration:

```bash
pip install numpy open3d pillow
cd single_image
```

**1. Point cloud only** (no window, writes files):

```bash
python lift_3d_pcd.py --scene 3025 --save --no-viz --no-gt --no-boxes
```

Writes to `out/lift_3d_pcd/003025/`: `points_world.ply`, `points_world.npz`
(`xyz`, `rgb`, `uv`, `xyz_cam`), `meta_world.json`.

**2. Visualize everything** (cloud + GT meshes + 3D boxes + camera frustum):

```bash
python lift_3d_pcd.py --scene 3025
```

Drag = rotate, scroll = zoom, shift+drag = pan, `q` = quit.

**3. All scenes, foreground only** (no window, object pixels only — uses `bgdepth`):

```bash
python lift_3d_pcd.py --scene all --mask fg --save --no-viz --no-gt --no-boxes
```

One folder per scene under `out/lift_3d_pcd/` (`003025/`, `003084/`, …).

Run `python lift_3d_pcd.py --help` for the remaining flags.

## License

Derived from **3D-FRONT** / **3D-FUTURE** (Alibaba); conventions cross-checked against
[Gen3DSR](https://github.com/AndreeaDogaru/Gen3DSR). Use is subject to the original
3D-FRONT license terms (non-commercial research) — please also cite the 3D-FRONT papers.