Structure Estimation from a Single View

Estimating contituent elements of a structure requires estimation of its position and orientation. We perform gradient descent over the position and orientation parameters of the structure minimizing the image distance between the projected elements of a 3D grid and the prominent linear edges derived from an image of the structure.

We can use vision to disambiguate language and language to disambiguate vision. Estimating structure from vision alone yields errors, even with multiple views. Addition of a lingustic description of the structure to the constraints from a single view can disambiguate the structure and produce a correct estimate.

Raw Detector Output Detector with Only Log-Segment Grammar Detector with Only Log-End Grammar Detector with Both Grammars Estimated Structure

Back to the Research page