3DIEBench-OOD / README.md
Hafez's picture
Update README.md
dcee90c verified

3DIEBench-OOD

Out-of-distribution (OOD) test set for 3DIEBench, designed to evaluate equivariance generalization to unseen rotation ranges.

Dataset Description

3DIEBench-OOD contains rendered views of ShapeNet objects with rotations sampled from ranges not seen during training:

Training (3DIEBench) OOD Test (This Dataset)
Rotation Range [-π/2, π/2] [-π, -π/2] ∪ [π/2, π]

This allows testing whether models can extrapolate equivariant predictions to OOD rotation angles.

For generation details, see data/3DIEBench/data_generate_ood.py in the seq-JEPA repository.

Structure

File Description
{synset}/{obj}/image_{i}.jpg Rendered image (256×256)
{synset}/{obj}/latent_{i}.npy 7D latent vector: [yaw, pitch, roll, floor_hue, spot_θ, spot_φ, spot_hue]

Usage

import numpy as np from PIL import Image

img = Image.open('04330267/8ff18f81de484792f0b94599b4efe81/image_2.jpg') latent = np.load('04330267/8ff18f81de484792f0b94599b4efe81/latent_2.npy')

latent = [yaw, pitch, roll, floor_hue, spot_theta, spot_phi, spot_hue]## Related Resources

Citation

If you use this dataset, please cite:

@inproceedings{ghaemi2025seqjepa, title={seq-{JEPA}: Autoregressive Predictive Learning of Invariant-Equivariant World Models}, author={Ghaemi, Hafez and Muller, Eilif Benjamin and Bakhtiari, Shahab}, booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}, year={2025}, url={https://openreview.net/forum?id=GKt3VRaCU1} }

@inproceedings{garrido2024sie, title={Self-supervised learning of split invariant equivariant representations}, author={Garrido, Quentin and Balestriero, Randall and Najman, Laurent and LeCun, Yann}, booktitle={International Conference on Machine Learning}, year={2024} }