original/README.md

metadata

pretty_name: AmbiSem-3D
tags:
  - 3d
  - image
  - amodal-3d-generation
  - semantic-ambiguity

AmbiSem-3D

A diagnostic set of 21 single-image cases for amodal 3D generation under semantic ambiguity, where the visible evidence in an image is ambiguous, consistent with multiple plausible object identities. Each sample provides an observation image, a category text label disambiguating the intended interpretation, and (for some samples) an object mask.

Folder layout

AmbiSem-3D/
├── manifest.json           # 21 sample entries
└── assets/
    └── <sample_id>/
        ├── input.png       # observation image
        └── mask.png        # optional, only for samples that include one

Manifest fields

Each entry contains:

Field	Description
id	Stable sample identifier matching the assets/<sample_id>/ folder
image	Observation image (relative to dataset root)
obs_object_image	Same as image for this dataset
prior_text	Disambiguating text caption
mask	Object mask, included only when available

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

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
    └── ...

python prepare_manifest_with_priors.py \
  --manifest path/to/AmbiSem-3D/manifest.json \
  --priors-root path/to/my_priors \
  --output path/to/AmbiSem-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 AmbiSem-3D or RelaxFlow, please cite:

@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}
}