<< Next >>

RESULTS - Direct Structure from Motion
Separation Bar
Direct Pose Registration - Structure from Motion (DSFM) algorithm computes directly the pose of the camera without the necessity to recover the inter-frame motion. The structure of the scene is formed by sets of 3D vertices characterized by affine invariant local image descriptors. In this way, by associating image patches extracted from camera views with the 3D vertices, we can recover the camera pose with respect to the scene model. In DSFM, the camera pose is obtained using a novel dual approach, allowing accurate camera pose estimations even in the presence of planar scenes, where most 3D reconstruction algorithms would fail.

Subsequently, the obtained camera poses are used to update the scene model as new features are tracked. Both camera pose estimation and scene model update steps use robust methods thus reducing the impact of poor camera pose/vertex estimations.

DSFM Flowchart
Fig 1. Flowchart of the DSFM algorithm, emphasizing the two stages.

DSFM algorithm works in two stages, as shown in Figure 1. First, it uses motion estimation techniques in order to obtain an initial model corresponding to a small sub-region of the scene. In the second stage, using the initial model as a "seed", the subsequent camera poses are computed by registering 2D features with 3D vertices in the scene model. For each newly acquired image, once the camera pose is recovered, the scene model is updated by adding vertices corresponding to newly tracked features. In this way, as the camera moves, the model is extended to represent new regions of the scene.

Figure 2 illustrates the entire recosntruction process for an underwater image sequence.

Fig 2. 3D modelling and ortho-mosaicing using DSFM.

As the data is being processed sequentially, camera pose and scene model estimations are constantly available, enabling the use of DSFM for online applications such as robot navigation and mapping, in situ scientific studies, etc. Figure 3 illustrates a series results using DSFM for both outdoor and underwater applications, including an example obtained in the presence of very challenging imaging conditions.

Fig 3. Some results of applying DSFM for different scenes, in both outdoor and underwater environments.