Automatic 3D modeling from range images

Many computer vision and robotics applications call for accurate three-dimensional (3D) models of real-world objects. Current 3D modeling techniques require significant manual assistance or make assumptions about the scene characteristics or data collection procedure. In this thesis work, we propose to fully automate the 3D modeling process without resorting to these restrictive assumptions. Given a set of unordered range images and no additional a priori information about the scene, our system will generate an accurate 3D reconstruction. Specifically, it is not necessary to know the relative pose between viewpoints or to indicate which views contain overlapping scene regions.

Our proposed automatic modeling system selects pairs of views that are likely to match and attempts to register them. The results are verified for consistency, but some incorrect matches may be locally undetectable and some correct matches may be missed. One of several discrete optimization techniques is employed to combine these potentially faulty pair-wise matches into a network of views called the model graph. Incorrect pair-wise matches are detected by the inconsistencies they produce elsewhere in the model graph, while missed matches are recovered by inferring new links in the graph between overlapping views. The overall model quality is improved by simultaneously registering all views before they are integrated together to form the final model. We demonstrate the utility of automatic modeling with an application called handheld modeling, in which a 3D model is automatically created from an object held in a person's hand.

View the proposal talk or download the proposal document (638K pdf).

Movies from the proposal talk:

• Data collection (2.9MB avi) - An example of the simplicity of data collection for handheld modeling. Just hold the object while the 3D scanner takes range images.
• Good match (1.5MB avi) - The result of uninformed pair-wise registration for a successful match.
• Bad match (1.4MB avi) - When registration fails, it can fail in surprising ways. In this case, the round head causes the two surfaces to match well in the region where they overlap.
• Initial angel model (2.7MB avi) - This is the model formed from correct pair-wise matches. Accumulating error leads to gaps in the model, which is the motivation for multi-view registration.
• Final angel model (2.3MB avi) - The final model after multi-view registration using all overlapping surfaces.