~ Research ~

[ 3D reconstruction from two rectified images ]

Here is an example of what can be done with two rectified images. With two uncalibrated images, you can only obtain a projective reconstruction of the scene, so you have to self-calibrate the camera. However, the case of rectified images, corresponding to a pure translation camera movement, is a critical case of self-calibration. Even with usual hypotheses on the camera internal parameters, the focal length remains undeterminable. So we have to tune the focal by hand.

murg

murd


movie (divx)

Once the scene has been reconstructed, we can play with the 3D model and insert objects:


movie (divx)

[ Plane-Based Self-Calibration: Guaranteed Solutions ]

The Plane-Based Self-Calibration problem is a non linear problem which consists of the estimation of the internal parameters of the camera along with the Euclidean structure of the plane. What we have at our disposal are several views of a plane and the inter-view homographies. We have shown that there is an interdependence in the parameterization introduced by Triggs in 1998.
Moreover, we have used Interval Analysis-Based Global Optimization to solve the problem and find guaranteed solutions to this problem.

From the resulting parameters, we can compute an Euclidean rectification of the key view and propagate this rectification to the other views in order to make a mosaic. The following illustration shows a rectified mosaic and 4 views among the 5 used to make the mosaic.

Matlab code: Download section.

     

murg

murg

murg

             

murg

murg

murg

[ Stereovision: Generation of stereo images with ground truth ]

  • Synthetic images with ground truth.

    Exact ground truth is only available for synhetic images. We proposed new photorealistic stereo images with the associated ground truth.
    Two examples:

    murg

    murd

    murt

    planteg

    planted

    plantet

    Blender script for the generation of ground truth from a 3d model: Download section.

    References :
    Benoît Bocquillon, Sylvie Chambon, Alain Crouzil. Segmentation semi-automatique en plans pour la génération de cartes denses de disparités. ORASIS05.

  • Ground truth from piecewise planar scenes.

    We considered piecewise planar scenes and we proposed a method and a tool to get semiautomatically a plane-based segmentation of the images and to generate accurate disparity maps. The semiautomatic segmentation helps to handle better the occlusions. The disparities are obtained from inter-images homographies.
    Here is a new stereo image pair with the associated ground truth:

    livresg

    livresd

    livrest

    Tool for the generation of ground truth from stereo images of a piecewise planar scene: Download section.

    References :
    Benoît Bocquillon, Sylvie Chambon, Alain Crouzil. Segmentation semi-automatique en plans pour la génération de cartes denses de disparités. ORASIS05.