Distance Estimation

[Emmanuelle Gouillart]

Hi Chris,

the short version is no, there is no function in scikit-image that can do
what the Stanford paper does. However, you can find some useful building
blocks in scikit-image and scikit-learn, that you can put together to
build your own version of the algorithm (from what I figured out after
a quick reading of
http://www.cs.cornell.edu/~asaxena/reconstruction3d/saxena_iccv_3drr07_learning3d.pdf)

- Felzenswalb superpixels are available in skimage.segmentation
  (http://scikit-image.org/docs/dev/auto_examples/segmentation/plot_segmentations.html)
- Local features can be computed using several functions of scikit-image,
  such as color transforms (color.ycbcr) and edge filters
(filters.sobel_h, filters.sobel_v, ...)
- A logistic regression algorithm is available in scikit-learn
  (http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html)

The trickiest part to code would be the MAP inference on the MRF, where I
guess you would have to code it yourself (I think).

So, it is feasible to assemble these different blocks, but it's not a
small project either.

As for other libraries, I'm afraid nothing comes to my mind. OpenCV does
have a depth estimation function, but only from stereo images.

Hope this helps,
Emma


On Wed, Aug 31, 2016 at 09:30:20PM -0400, Chris Spencer wrote:
> Essentially, what's described here:

> http://www.cs.cornell.edu/~asaxena/learningdepth/

> On Wed, Aug 31, 2016 at 9:29 PM, Chris Spencer <chrisspen at gmail.com> wrote:

>     Given a JPG image of a scene, convert that into a depth map, where each
>     pixel is associated with a discrete distance estimate.

>     On Wed, Aug 31, 2016 at 7:12 PM, Stefan van der Walt <stefanv at berkeley.edu>
>     wrote:

>         Hi Chris

>         On Sun, Aug 28, 2016, at 17:31, Chris Spencer wrote:
>         > Is there any algorithm in scikit-image that I could use to train a
>         predictor that could estimate the distance at each pixel in an image?

>         > I've gone through all the examples, and none of them seemed to
>         directly apply. The image segmentation example seemed to be the
>         closest, but that didn't use supervised learning, so I wasn't sure how
>         I could adapt it's code.

>         Could you explain what you mean by "estimate the distance"?

>         Stéfan