6.344  Digital Image Processing

Course Wrap-Up

Hi All,

It was great to see all your project presentations last week. I hope you will all be able to use the skills you have learned from this course in your future endeavors.

I also wanted to thank Tien-Ju for all of his help with the course this year! He was instrumental in helping to add new material on CNNs into the problem sets this year (amongst his many other TA duties).

Finally, if you haven't done so already, please submit your evaluations at http://web.mit.edu/subjectevaluation

The website closes Monday morning at 9AM. Your feedback is highly valued by the department (resource allocation), faculty (course development) and your fellow students (HKN underground guide).

Have a great summer!

Vivienne

Announced on 21 May 2017  11:18  a.m. by Vivienne Sze

Final presentations today | Please arrive on time

Final presentations are today!

*** As a courtesy to your fellow students, please arrive to class on time, as we must start at 11:05am sharp in order to fit everyone in.***
There will be a small prize for best presentation as voted by the audience (so audience members, please arrive on time).

Finally, if you haven't done so already, please submit your evaluations at http://web.mit.edu/subjectevaluation
[we are currently at 27% response rate]
Your feedback is highly valued by the department (resource allocation), faculty (course development) and your fellow students (HKN underground guide).

Announced on 16 May 2017  8:04  a.m. by Vivienne Sze

Final Presentations on Tuesday May 16

Hi All,

For the final presentations please plan on
* 8 minutes for a 1 person team
* 10 minutes for a 2 person team
* 12 minutes for a 3 person team

We'll also have 1 to 2 minutes for questions.

*** As a courtesy to your fellow students, please arrive to class on time, as we must start at 11:05am sharp in order to fit everyone in.***
Also, there will be a small prize for best presentation as voted by the audience (so audience members, please arrive on time).

We have many exciting projects this year, covering a variety of topics and applications. Tien-Ju and I are very much looking forward to seeing all your presentations.

Announced on 11 May 2017  1:19  p.m. by Vivienne Sze

Project Deadline Extension and more...

Hi All,

A couple quick announcements:
* The project report deadline has been extended to Sunday, May 14 @ 10PM (slides are still due May 15 @ 5PM). *All presentations will be on May 16*
* All problem sets have been graded; please contact Tien-Ju ASAP if you have any questions
* Congrats to Diana Wofk and Tiffany Le for training the best performing denoising CNN models on both Gaussian and salt-and-pepper noise! They received the bonus points award on Pset #6.
* The course evaluation website is up: http://web.mit.edu/subjectevaluation/evaluate.html
Please take a few moments to provide feedback on the course.

See you all Thursday!

Announced on 09 May 2017  9:42  p.m. by Vivienne Sze

Clarification on the question asked in the class

Clarification on the question asked in the class: Why is the score on the right hand side plot constant with alpha in case of the step edge?

The plots in the slides were from this paper: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=5206815
The 1D sharp and blurred graphs are plotted on an axis going from 0 to 1. So the value of the gradient at the step edge is 1, so raising to any power (alpha) will result in 1.

On the right hand side plots, a lower score is better. For step edges, MAPx,k usually succeeds. While the Gaussian prior favors the blurry explanation, appropriate sparse priors (α < 1) favor the correct sharp explanation. For the narrow peaks the sharp explanation is favored only for low alpha values. However, the sparse models describing natural images are not binary, they are usually in the range α ∈ [0.5, 0.8].

Natural images have more narrow peaks than step edges so MAPx,k approach usually favors the delta kernel solution.

Announced on 02 May 2017  1:10  p.m. by Tien-Ju Yang