Adaptation of Synthetic Data for Coarse-to-Fine Viewpoint Refinement

Pau Panareda Busto, Joerg Liebelt and Juergen Gall

Abstract

The quality of learning-based pose estimation still heavily relies on manual training data annotations. However, the manual labeling of large datasets is costly and frequently limited to a few coarse viewpoint annotations of varying accuracy. In this work, we propose to refine such coarse pose annotations with a domain adaptation approach, where the source domain consists of fine-grained pose annotations generated from synthetic computer graphics models, and the target domain of coarse manual pose annotations of a real dataset. Our domain adaptation step computes a linear map which aligns corresponding samples from the two domains and allows for the refinement of the manual pose labels using the transformed synthetic ones. Experiments show that we significantly improve pose estimation on several state-of-the-art datasets.

Session

Poster 1

Files

PDF iconExtended Abstract (PDF, 892K)
PDF iconPaper (PDF, 5M)

DOI

10.5244/C.29.14
https://dx.doi.org/10.5244/C.29.14

Citation

Pau Panareda Busto, Joerg Liebelt and Juergen Gall. Adaptation of Synthetic Data for Coarse-to-Fine Viewpoint Refinement. In Xianghua Xie, Mark W. Jones, and Gary K. L. Tam, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 14.1-14.12. BMVA Press, September 2015.

Bibtex

@inproceedings{BMVC2015_14,
	title={Adaptation of Synthetic Data for Coarse-to-Fine Viewpoint Refinement},
	author={Pau Panareda Busto and Joerg Liebelt and Juergen Gall},
	year={2015},
	month={September},
	pages={14.1-14.12},
	articleno={14},
	numpages={12},
	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
	publisher={BMVA Press},
	editor={Xianghua Xie, Mark W. Jones, and Gary K. L. Tam},
	doi={10.5244/C.29.14},
	isbn={1-901725-53-7},
	url={https://dx.doi.org/10.5244/C.29.14}
}