Description
General Information
Lecture 2, Nov 10 (Michal): Fourier (2/2)
Lecture 3, Nov 17 (Ronen): Geometry (1/4)
Lecture 4, Nov 24 (Ronen): Geometry (2/4)
Lecture 5, Dec 1 (Shai): Deep Learning (1/3)
Lecture 6, Dec 8 (Shai): Deep Learning (2/3)
Lecture 7, Dec 15 (Shai): Deep Learning (3/3)
Lecture 8, Dec 22 (Assaf): Deep Generative Models
*Dec 29 – FGS Hannuka vacation (NO CLASS)*
Lecture 9, Jan 5 (Michal): Motion & Video (1/2)
Lecture 10, Jan 12 (Michal): Motion & Video (2/2)
Lecture 11, Jan 19 (Ronen): Geometry (3/4)
Lecture 12, Jan 26 (Ronen): Geometry (4/4)
Lecture 13, Feb 2 (Shimon): Recognition
Final exam: 60%
* Secondary: March 19, 2020, 10:00-13:00 (Ziskind, Rm 1)
Announcements
Hi,
Due to coronavirus we need to postpone Moed B exam that was scheduled to this Thursday (Mar 19th).
There will be no exam on Mar 19th.
A new date will be announced once things settles down.
Wishing you all good health,
Dear students,
Moed A exams have been graded.
You may review your notebooks which are found at Shai's office (221).
Best,
The Course Staff.
Hi,
I added a summary I wrote about geometry.
It doesn't include all the material, but I hope it would help if you are still confused about projective geometry, two-view geometry, etc.
Note that I wrote it myself and Ronen didn't see it yet, so it's not really official course material (and please let me know if you find any errors).
You can view it on the course page: https://piazza.com/weizmann_institute_of_science/winter2020/20204011/resources
Shira
Hi,
Exercise 5 has been checked.
You can find your checked solution in your personal Dropbox folder.
Good luck in the exam!
Best,
The Course Staff
Dear students,
Here are some details about the exam:
- The exam is a written in-class examination.
- The exam is closed book. You are not allowed to use any study material during the exam.
- The duration of the exam is 3 hours.
- The exam contains 3-4 questions. You should answer all the questions.
- The material covered in the exam is outlined in the Resources section. It includes:
- Image Processing
- Geometry
- Deep Learning
- Recognition
- Please find posted exams from past years to assist in preparation.
Good luck !
The Course Staff
Dear students,
Please be noted that the deadline was extended further to this Tuesday (Feb 11, until the end of the day).
Best,
The Course Staff
Dear students,
Exercise 4 has been graded.
You can find your checked solution in your personal Dropbox folders.
Please consider a few general remarks regarding the exercise:
- The number of iterations is not set in stone. It is contingent upon the type of optimizer, the learning rate, your regularization, your model, the particular image, etc. Many of you had set the max_iteration count as a fixed number across experiments, say 2500, disregarding the resulting convergence stage of your training at that finishing line. A complete view should show reaching a top validation PSNR (or minimal validation loss), followed by overfit (or plateau), while training loss continues to decrease. If your train and validation losses are both still decreasing at your end-of-training, this means that your training is not fully exhausted.
- Overfit is a phenomenon where your validation loss is compromised, namely starts to increase, due to over-fitting of the model to your training data, namely training loss continues to decrease. Overfit is NOT defined as when train/validation losses cross.
- Note that PSNR is defined as the inverse of MSE, up to a logarithm. Therefore minimizing MSE is directly maximizing the PSNR.
- ADAM is a modern variant of SGD with adaptive learning rate. However they are not comparable at the same (initial) learning rate. The learning rates for SGD are usually 100-1000x larger than those used for ADAM, at comparable performance.
- This is a computer vision course. It is highly important to include high quality plots and images in submitted reports/papers, especially when dealing with a task like denoising. This includes avoiding suffering artifacts from low quality and poor resolution snippets, and to conduct proper side-by-side comparisons as some of you did. Presenting side-by-side the ground-truth, the noisy and the denoised images provides a comprehensive view for visual inspection. Every result should be contrasted against its baseline at similar scales, be it a curve or presented images.
Best,
The Course Staff
Dear students,
The submission deadline for Ex5 is now extended to Sunday Feb 9 2020 (until the end of the day).
Best,
The Course Staff
Name | Office Hours | |
---|---|---|
Ben Feinstein | When? Where? | |
Guy Gaziv | When? Where? | |
Shira Kritchman | When? Where? | |
Michal Irani | When? Where? | |
Shai Bagon | When? Where? | |
Assaf Shocher | When? Where? | |
Ronen | When? Where? | |
Ganit Kupershmidt | When? Where? |