This page supplements our paper on multiscale atlas estimation.
Atlas estimation on digit images from the MNIST handwritten digits database
Below are displayed movies showing the estimation of template images using either the original algorithm, either our coarse-to-fine strategy, with varying numbers of control points. Template images are displayed every 10 iterations until convergence. In all cases, atlas estimation was performed using the same set of 20 images. ācpā stands for control points.
| Number of cp |
|
|
|---|---|---|
| 49 |
|
 |
| 100 |
 |
 |
| 196 |
 |
 |
| 361 |
 |
 |
Template image and initial vector field estimation:
Below is displayed the estimation of the initial momenta that wrap the template image to the image of a subject. The magnitude of the vectors was multiplied by a factor of 2 or 3 for visualization purposes.
| Number of cp |
Subject image | |
|
|---|---|---|---|
| 49 |
|
|
 |
| 100 |
|
|
 |
Atlas estimation on images of artificial characters
| Number of cp |
||
|---|---|---|
| 36 |
 |
 |
| 49 |
 |
 |
| 100 |
 |
 |
| 196 |
 |
 |