Learning 2D Invariant Affordance Knowledge for 3D Affordance Grounding

AAAI 2025

Motivation

Existing image-guided 3D affordance grounding methods rely on a single human-object interaction image and can overfit to its appearance and geometry. MIFAG instead extracts the interaction patterns shared by multiple images of the same affordance, stores them as invariant affordance knowledge, and adaptively fuses this knowledge with point-cloud features for more generalizable 3D affordance grounding.

Setup

pip install -r requirements.txt

# Download the ImageNet-pretrained ResNet-18 weights used for training
mkdir -p pretrained
python -c "import torch; from torchvision.models import resnet18, ResNet18_Weights; torch.save(resnet18(weights=ResNet18_Weights.DEFAULT).state_dict(), 'pretrained/resnet18.pth')"

To use another ResNet-18 checkpoint, pass its path with --resnet18_path.

Data Preparation

Download the PIAD dataset by following the IAGNet data download instructions. Place the downloaded images and point clouds under Data/Seen/ and Data/Unseen/. The MIPA split files used by this project are provided in data/mipa_splits/.

Training

# Seen setting
bash scripts/train_seen.sh

# Unseen setting
bash scripts/train_unseen.sh

For a custom dataset root, append --data_root /path/to/mipa_root. Checkpoints are saved under runs/train/.

Evaluation

Download the pretrained checkpoints and evaluate the seen or unseen setting with:

bash scripts/eval_seen.sh /path/to/mifag_mipa_seen.pt
bash scripts/eval_unseen.sh /path/to/mifag_mipa_unseen.pt

The evaluation reports AUC, aIOU, SIM, and MAE.

Pretrained Checkpoints

Both checkpoints use the full MIFAG model (IAM + ADM) with two reference images. The seen model uses four IAM layers, while the unseen model uses six IAM layers.

Setting	Checkpoint	AUC	aIOU	SIM	MAE
Seen	Download	0.8510	0.2050	0.5681	0.0912
Unseen	Download	0.7113	0.0523	0.3148	0.1358

Acknowledgements

This codebase is built upon the official IAGNet implementation. We thank the authors for releasing their code and the PIAD dataset.

Citation

If you find this work useful, please cite:

@inproceedings{gao2025mifag,
  title={Learning 2d invariant affordance knowledge for 3d affordance grounding},
  author={Gao, Xianqiang and Zhang, Pingrui and Qu, Delin and Wang, Dong and Wang, Zhigang and Ding, Yan and Zhao, Bin},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={3},
  pages={3095--3103},
  year={2025}
}