Ping TAN
Associate Professor
School of Computing Science
Simon Fraser University

Structure-from-Motion & Visual SLAM
We are working to improve the efficiency and accuracy of the structure-from-motion algorithm, and to apply it for simultaneous localization and mapping (SLAM) applications.

Structure-from-Motion (SFM)
Incremental SFM reconstructs images one by one, which is sub-optimal and suffers from larger drifting error. We develop algorithms to solve all cameras simultaneously in a global fashion based on a novel geometric error for camera registration (ICCV 2013, BMVC 2015).
Ambiguous feature matching leads to incorrect 3D reconstructions. We designed an algorithm to upgrade an incorrect reconstruction to a correct one (CVPR 2012).

Visual SLAM
Many visual SLAM algorithms focus on the scenario of a single robot working in a static environment. However, our real world is full of moving objects. Furthermore, with the development of microrobots, we can easily have multiple robots working collaboratively. We develop the first visual SLAM algorithm for multiple independently moving robots in dynamic environments (PAMI 2012). Video demos can be found at the project page.
This work has been integrated with ground robots (RSS Workshop 2015). A video demo is at the project page (fun to watch, two robots chasing and orbiting a teddy bear :) ).

Computational Videography

Mobile devices make video capturing easy. We develop algorithms to enhance consumer videos by reducing the shake, stabilizing white balance changes, defogging, and generating artistic shots.

Video Stabilization
Videos captured from handheld devices are often very shaky. We propose a new motion model, i.e. Bundled Camera Path, and a novel filtering technique for stabilization (Siggraph 2013). This filtering technique is further generalized to smooth optical flows (CVPR 2014).
A depth camera simplifies the stabilization problem (CVPR 2012)

Tracking shots create expressive artistic photographs, where the moving object is sharp and the static background is blurred. Tracking shots are hard to capture, even for professionals. We develop an algorithm to automatically generate realistic, 3D-aware tracking shots from consumer videos. (Siggraph Asia 2014).

Image-based Modeling
Realistic 3D models are important for computer graphics applications. Most existing modeling systems still heavily rely on user interactions to manually model details. We are working towards the goal of automatic creation of realistic 3D models from images of real objects.
Image-based Tree Modeling
Trees are everywhere and are difficult to model in a realistic way. We developed methods to easily recover realistic 3D models from 2D images. This approach was applied to small plants (SIGGRAPH 2006) and trees (SIGGRAPH 2007). We also developed a method for the extreme case when only one input image is available (SIGGRAPH Asia 2008). Video demos and 3D models can be found at the project page.
Image-based Architecture Modeling
There is a strong demand for the photo-realistic modeling of cities for games, movies and map services. We applied the image-based modeling approach with street-level images to model buildings at the scale of a city block (SIGGRAPH Asia 2008) or a single building (SIGGRAPH Asia 2009). Video demos and 3D models can be found at the project page.
Data-driven Media Synthesis
While professional artists can quickly create high-quality images and animations, most of the ordinary people cannot do that. We are exploiting the large amount of visual data to design algorithms in a data-driven approach to simplify media creation.
A picture is worth a thousand words, and not surprisingly, people often use pictures and animations to convey ideas and stories. We developed methods to automatically convert an annotated freehand doodle into a realistic picture (SIGGRAPH Asia 2009) or a comic strip (TVCG 2012). More details can be found in the paper and the project page. The image search and filtering technique behind this application is extended in the 'semantic colorization' (SIGGRAPH Asia 2011) to allow users quickly to turn a grayscale image into color.

Virtual Clothes Try-on
Virtual clothes fitting applications allow users to see how they appear in chosen clothes without physically putting them on. We have taken an image-based approach for virtual clothes try-on. From a database of a segmented garment undergoing various motions, we synthesize arbitrary animations of the garment by rearranging the temporal order of database frames. More details can be found in the paper (SIGGRAPH Asia 2012 Technique Briefs) and the project website.

Reflectance Analysis
An image of an object is determined through complex interactions between its shape, material, illumination, and the imaging process. We are seeking to invert this process and recover scene information.

Photometric Stereo
Surface normal directions can be precisely determined from multiple images captured by a fixed camera under variant illumination conditions.
Autocalibration: When illumination conditions are unknown, there is an intrinsic shape/lighting ambiguity. We showed this ambiguity can be solved by the symmetry (isotropy and reciprocity) of BRDFs. The analysis can be performed on the Gaussian sphere (CVPR2007) or the real projective plane (CVPR2009). For a more comprehensive summary, please refer to the journal version (PAMI2011). We also developed auto-calibration algorithms based on partial information in surface albedos (CVPR2010) and lighting configurations (ECCV2010).
Super-resolution: We can reconstruct the subpixel level geometric structures on rough surfaces (ECCV2006) (PAMI2008) by applying our multi-resolution reflectance model (EGSR2005) (TVCG2008).
Complex material: We also studied photometric stereo with general non-Lambertian surfaces with a novel bi-quadratic reflectance model (CVPR2012) and the analysis of BRDF monotonicity (ECCV2012).

Reflectance Separation
Intrinsic images: An image might be separated into two intrinsic components: texture and shading. We introduce a non-local constraint to relate texture values at far apart pixels (CVPR2008). Closed-form solution can be derived from this constraint (PAMI2012). Depth sensor can also be used to further improve this separation (ECCV2012).
Highlight removal:
Surface reflectance can also be separated into diffuse and specular components, each with different physical properties. We developed methods to achieve this separation of regular surfaces (ICCV2003) and texture surfaces (CVPR2006).

Last modified in Jul 2015