File size: 3,247 Bytes
684a387
 
 
 
 
 
 
 
 
 
 
2ee32c1
c5e787e
 
2ee32c1
 
 
 
 
 
ff1906a
c5e787e
684a387
 
 
 
 
 
 
 
 
 
 
 
2ee32c1
684a387
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2ee32c1
 
 
 
 
 
 
684a387
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
pretty_name: "ExtremeOcc-3D"
tags:
- 3d
- image
- amodal-3d-generation
- extreme-occlusion
---

# ExtremeOcc-3D

ExtremeOcc-3D is a benchmark for amodal 3D generation under severe object
occlusion, introduced with [**RelaxFlow: Text-Driven Amodal 3D
Generation**](https://arxiv.org/abs/2603.05425), an ICML 2026 Spotlight paper.

The benchmark contains 264 single-view indoor scenes. Each sample provides an
occluded scene render, a target-object mask, an isolated render of the visible
object part, a category text label, and ground-truth 3D assets for evaluation.

Use `manifest.json` as the entry point. All paths in the manifest are relative
to this dataset root. Code and runners are available in the
[**RelaxFlow repository**](https://github.com/viridityzhu/RelaxFlow).

## Folder layout

```
ExtremeOcc-3D/
├── manifest.json           # 264 sample entries
├── assets/                 # observation images and masks per sample
│   └── <scene>/<view>/{rendered_scene.png, occluded_mask.png, rendered.png, ...}
└── gt_assets/              # ground-truth assets shared by views of the same object
    └── <object_hash>/{mesh.ply, view_0000.png, ...}
```

## Manifest

Each entry contains:

| Field | Description |
|---|---|
| `id` | Stable sample identifier |
| `image` | Occluded scene render (relative to dataset root) |
| `mask` | Object mask for the target object |
| `obs_object_image` | Isolated render of the observed (visible) part of the object |
| `prior_text` | Object category label |
| `gt_render_dir` | Directory of multi-view GT renders for the target object |
| `gt_mesh` | Ground-truth mesh (`.ply`) for the target object |

All paths are relative to this `ExtremeOcc-3D/` directory.

## Usage

Download this repository, then pass `manifest.json` to the RelaxFlow batch
runner. If you generate prior images separately, use
`prepare_manifest_with_priors.py` from the RelaxFlow release to attach them to a
copy of the manifest.

## Adding prior images

The manifest does not include a `prior_images` field. Generate priors with any image generator you prefer, organize them per sample, then run the helper to produce a complete manifest:

```
my_priors/
└── <sample_id>/             # must match the `id` field in manifest.json
    ├── prior_0.png
    ├── prior_1.png          # multiple priors are supported
    └── ...
```

```bash
python prepare_manifest_with_priors.py \
  --manifest path/to/ExtremeOcc-3D/manifest.json \
  --priors-root path/to/my_priors \
  --output path/to/ExtremeOcc-3D/manifest_with_priors.json
```

Accepted image extensions: `.png .jpg .jpeg .webp`. Files inside each `<sample_id>/` are picked up in sorted order. The output `manifest_with_priors.json` plugs directly into the relaxflow batch runners.

## Citation

If you use ExtremeOcc-3D or RelaxFlow, please cite:

```bibtex
@inproceedings{zhu2026relaxflow,
  title     = {RelaxFlow: Text-Driven Amodal 3D Generation},
  author    = {Zhu, Jiayin and Fu, Guoji and Liu, Xiaolu and He, Qiyuan and Li, Yicong and Yao, Angela},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2603.05425}
}
```