2018 Fiscal Year Annual Research Report
Photo-realistic Image-based Modeling and Rendering in the Wild
| Project/Area Number |
17F17350
|
|---|
| Research Institution | Osaka University |
|---|
Principal Investigator |
松下 康之 大阪大学, 情報科学研究科, 教授 (30756507)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
WAECHTER MICHAEL 大阪大学, 情報科学研究科, 外国人特別研究員
|
|---|
| Project Period (FY) |
2017-11-10 – 2020-03-31
|
| Keywords | コンピュータビジョン / 三次元形状復元 |
|---|
| Outline of Annual Research Achievements |
We continued our work on calibrating the positions of point light sources based on the observation of shadows of objects with unknown 3D positions. Our paper was accepted as an oral presentation at the European Conference of Computer Vision (an oral presentation is only awarded to less than 2% of all submissions), one of the top conferences in our field.
Following that, we worked on further investigating the mathematical connections between camera geometry and shadow projection geometry. We found that in multi-view settings, i.e., a static scene with a moving light source, the movement of all shadows can be explained by the same geometric and algebraic principles as the movement of image points in photos of a static scene with a moving camera: Fundamental matrices, trifocal tensors, and quadrifocal tensors. We showed how they are derived and how they can efficiently and stably be estimated from data.
We just submitted this work to the top journal in our field, the International Journal of Computer Vision and are positive that it gets accepted for publication. More importantly, we consider that this is an important fundamental advancement towards high-fidelity 3D reconstruction, which is crucial to our research goal.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
難関国際会議での登壇発表やトップジャーナルへの投稿など、順調に研究が進んでいる。
|
| Strategy for Future Research Activity |
For the remainder of my fellowship until the end of October 2019, we plan to
-investigate whether the problem of finding corresponding points in multi-view imagery is a convex optimization problem. This would mean that correspondence search can be done with gradient descent methods. In particular, we want to analyze the search for corresponding shadows under changing light positions based on our shadow geometry findings mentioned earlier.
-A consequence of our research on shadows mentioned earlier is, that the sun and the shadows that it casts effectively act as a camera that observes the world from many viewpoints and any planar ground can act as a “camera sensor”. We want to verify whether this can be used in practice for reconstructing the geometry of static objects in the world that stand on planar ground by simply taking several pictures of an outdoor scene with a fixed camera throughout one day.
|
