by M Wimmer and B Radig
Abstract:
Tracking 3D models in image sequences essentially requires determining their initial position and orientation. Our previous work identifies the objective function as a crucial component for fitting 2D models to images. We state preferable properties of these functions and we propose to learn such a function from annotated example images. This paper extends this approach by making it appropriate to also fit 3D models to images. The correctly fitted model represents the initial pose for model tracking. However, this extension induces nontrivial challenges such as out-of-plane rotations and self occlusion, which cause large variation to the models surface visible in the image. We solve this issue by connecting the input features of the objective function directly to the model. Furthermore, sequentially executing objective functions specifically learned for different displacements from the correct positions yields highly accurate objective values.
Reference:
Initial Pose Estimation for 3D Models Using Learned Objective Functions (M Wimmer and B Radig), In Proceedings of the 8th Asian Conference on Computer Vision (ACCV07) (Y Yagi, SB Kang, IS Kweon, H Zha, eds.), Springer, volume 4844, 2007.
Bibtex Entry:
@inproceedings{wimmer_initial_2007,
author = {M Wimmer and B Radig},
title = {Initial Pose Estimation for {3D} Models Using Learned Objective Functions},
booktitle = {Proceedings of the 8th Asian Conference on Computer Vision ({ACCV07)}},
year = {2007},
editor = {Yagi, Yasushi and Kang, Sing Bing and Kweon, In So and Zha, Hongbin},
volume = {4844},
series = {{LNCS}},
pages = {332--341},
address = {Heidelberg},
month = {nov},
publisher = {Springer},
abstract = {Tracking {3D} models in image sequences essentially requires determining
their initial position and orientation. Our previous work identifies
the objective function as a crucial component for fitting {2D} models
to images. We state preferable properties of these functions and
we propose to learn such a function from annotated example images.
This paper extends this approach by making it appropriate to also
fit {3D} models to images. The correctly fitted model represents
the initial pose for model tracking. However, this extension induces
nontrivial challenges such as out-of-plane rotations and self occlusion,
which cause large variation to the models surface visible in the
image. We solve this issue by connecting the input features of the
objective function directly to the model. Furthermore, sequentially
executing objective functions specifically learned for different
displacements from the correct positions yields highly accurate objective
values.},
isbn = {978-3-540-76389-5},
}