The famous landmarks and tourism sites around the world, such as the Rome Trevi Fountain and the Tokyo Train Station, have been photographed hundreds of thousands of times, at different time and places by different photographers.
Nowadays, as the online photo and video sharing sites, like Flickr and YouTube, are becoming popular, a portion of these images are publicly available on the Internet. These Internet image collections are tremendously huge in quantity and extremely diverse in viewpoints, color tones, light/shadow effects and so on. If the embedded viewpoint and appearance variations could be properly exploited, large-scale Internet image collections are capable of offering us abundant in-depth information, including 3D geometric structure, surface reflective properties, texture and more, which are not available in small scattered image sets.
Our first-hand experiences indicate that traditional models and inference algorithms are inadequate for large-scale Internet image collections, due to their poor accuracy or (and) unfavorable computational efficiency. Therefore, the primary research topics include:
- Incremental 3D reconstruction pipeline without using intermediate 3D information;
- Fast minimal problem solvers for the relative and absolute pose estimation;
- Reflectance and texture recovery via intrinsic image decomposition guided by sparse correspondence.
- Photo tourism with user designations
- Mobile image-based localization
- Posterior high-fidelity image editing
鄭 銀強［コンテンツ科学研究系 助教］