Combining Discriminative and Generative Methods for 3D Deformable Surface and Articulated Pose Reconstruction
| Mathieu Salzmann | Raquel Urtasun |
| EECS & ICSI, UC Berkeley | TTI, Chicago |
The following videos are encoded in MPEG-4, and were tested with QuickTime Player and VLC Player.
Deformable Surfaces
Additional results comparing our approach to optimizing the image likelihood using the discriminative predictor only for initialization. These results were computed using Pyramid HOG features as inputs. From left to right: Cardboard data with well-textured images, cardboard data with poorly textured images, similar results for inextensible meshes. Note that using constraints significantly reduces the error. Additionally, even when learning the basis of the representer theorem fails (third plot), constraining the prediction yields very accurate results. This shows that every step of our approach is important for accurate reconstruction.
Results on real images. In each video, we show on the top row, the recovered mesh reprojected on the original images, and on the bottom row the same mesh seen from a different viewpoint. From left to right, results correspond to the original GP, the constrained GP when optimizing y, the constrained GP used in conjunction with an image likelihood when optimizing y, and similar results when optimizing k*.
- paper.mp4: Recovering the shape of piece of paper.
- cloth.mp4: Reconstructing a deforming piece of cloth.
Articulated Poses
Videos demonstrating the performance of our approach on articulated pose estimation. Each video shows randomly selected samples from our test sets.
- hand.mp4: Estimating the pose of a hand from Pyramid HOG features.
- humaneva.mp4: Estimating full body poses from noisy 2D joint locations.
