BAT3R: Bootstrapping Articulated 3D Reconstruction from 2D Image Collections

Jakub Zadrożny, Oisin Mac Aodha, Hakan Bilen
University of Edinburgh
Accepted to ECCV 2026! 🎉
arXiv Poster Code (coming soon!)
[TL;DR] BAT3R leverages unannotated 2D image collections and a single rest-pose mesh to automatically generate synthetic paired 3D training data, achieving high-quality 3D reconstruction without expensive manual curation.
BAT3R Teaser Grid

BAT3R recovers challenging articulations, camera poses, and rest-pose correspondences across categories. It enables efficient mesh fitting for 3D reconstructions with topologically sound geometry.

Abstract

3D reconstruction of articulated objects from a single image is challenging because large training datasets with paired image and 3D supervision are difficult to obtain. Recent point map–based methods achieve strong performance but rely on synthetic datasets rendered from manually created articulated 3D assets with carefully curated pose distributions. While camera viewpoints can be easily sampled, generating realistic object articulations remains costly and labor-intensive. We propose a training framework that reduces this requirement by leveraging unannotated 2D images collections with only a single rigged canonical mesh per category.

Starting from a weak 3D shape predictor trained on canonical-pose renders, we iteratively estimate object articulation and camera pose by fitting the mesh to predicted point maps. The recovered articulations and viewpoints are then used to render updated synthetic training data, progressively improving the predictor. Despite using substantially weaker 3D supervision, our models achieve performance comparable with DualPM, which requires manually curated articulated training datasets.

Results with SAM-3D and Trellis.2

Qualitative comparisons to foundation models on real-world images across diverse categories.

Ablations

Visualizing point cloud predictions and extracted fitted meshes across ablation settings.

Method Overview

Method Pipeline

(Left) We render the canonical rigged mesh from sampled camera viewpoints to create an initial paired 2D--3D training set, which is used to train the initial 2D to 3D predictor. (Right) Given an image collection, the current predictor predicts posed and canonical point maps. We fit the canonical mesh to these predictions, recover an articulated mesh and camera viewpoint, and render the result to obtain new synthetic training pairs with exact point-map supervision. The predictor is then retrained on this generated dataset, yielding an improved model. The process is repeated for 4 iterations.

BibTeX

@misc{zadrozny2026bootstrapping,
      title={Bootstrapping Articulated 3D Reconstruction from 2D Image Collections}, 
      author={Jakub Zadrozny and Oisin Mac Aodha and Hakan Bilen},
      year={2026},
      eprint={2607.03891},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.03891}, 
}