I have been working on the three main steps of the modelling pipeline: camera motion and calibration, 3D reconstruction and texture mapping:
Image Acquisition
3D Reconstruction
Camera Calibration
Texture Mapping
Camera calibration
We propose a novel approach of camera calibration based on the concept of silhouette coherence. For a given set of silhouettes and corresponding camera matrices, the silhouette coherence measures how likely is that a 3D object might have generated those silhouettes using the camera projection matrices. Roughly speaking, we perform a comparison between a set of silhouettes, and the silhouettes of the reconstructed visual hull using the camera matrices. If both the silhouettes and the camera matrices are correct, then the input silhouettes and the silhouettes of the visual hull will be exactly the same. If the camera matrices or the silhouettes are noisy, the silhouettes of the visual hull will always be included in the original silhouettes. We propose a fast algorithm to measure the mismatch between both sets of silhouettes. Camera calibration is simply performed as a mismatch minimisation problem.
both silhouettes and projection matrices are correct
silhouettes or projection matrices are not correct
3D reconstruction
Textured objects...
Our current approach consists of exploiting feature points and silhouettes for the calibration step, and texture and silhouettes for the 3D modelling step.
Take a look at some 3D reconstructions I have done with Roberto: one of Antony Gormley's sculptures, one of the Parthenon marbles at the British museum and a Henry Moore sculpture.
Crouching sculpture by Antony Gormley
Original sequence sample
Horsemen from the west frieze of the Parthenon Greek, about 438-32 BC Acropolis, Athens, Greece
Original sequence sample
Henry Moore sculpture
Original sequence sample
If your browser supports java, click on the following gallery to visualize online some of the 3D reconstructions:
... and textureless objects
In recent work with George Vogiatzis we have obtained full 3D reconstructions of porcelain objects from 36 views under changing but unknown lighting. This work has two main contributions:
(a) We generalise photometric stereo in the multiple view case which allows for 360 degree closed surface reconstructions.
(b) We show how the direction and intensity of an infinite distance light-source can be obtained from the silhouettes of a Lambertian object.
Porcelain figurine
Original sequence sample
Porcelain vase, Fitzwilliam museum
Original sequence sample
For a more detailed look at some other models we reconstructed using this technique, click on the following gallery, all you need is a java enabled browser:
See my publication list for a more detailed explanation of the different algorithms. You can also download more than 100 reconstructions with their corresponding high quality texture map from the 3D archive.
