Recently I have been at IWCV — invitation-only, no paper publication workshop in Modena, Italy.

Group photo of IWCV participants, source http://iwcv2018.unimore.it/

According to organizers,

We have noticed that usefulness of CVPR for thoughtful ideas sharing and discussion becomes diminishing and started IWCV. And I have asked speakers to prepare not their usual talks

And it was cool. Lots of “not usual” talks from top-notch CV researchers were really inspiring. Venue was also great — Modena is >1000 years old city, where Enzo Ferrari started his famous company.

But lets go straight to interesting and relevant(to me!) talks.

  1. Bill Freeman presented “Looking to listen” talk. Main idea — image helps to recognize sounds and vice versa. There is even phenomena, when people in glasses started to hear better. Results were impressive…and rejected from CVPR. They resubmitted to SIGGRAPH and got oral.
Figure from https://arxiv.org/pdf/1804.03619.pdf

2. Neil Lawrence gave cool talk on uncertainty in deep learning. Especially, he insisted that one should use loss function in the direct sense of “what we pay for mistake” and shown have to add uncertainty into simple linear regression, so result started make sense. Here are some

Evolution is not survival of the fittest, it is non-survival of non-fitting…
Using loglikelyhood is just delaying decision on loss function…
Good thing is freedom minimization — when you have no other options other than correct…

3. Stephan Roth presented perception-inspired “Detail preserving pooling” (DPP) and way how to propagate variance in vanilla CNNs. DPP is generalization of average and max pooling, more general than GeM pooling :)

Figure from https://arxiv.org/pdf/1804.04076.pdf

4. Yoav Y. Schechner presented a way, how to decompose light sources in image by estimating phase of AC current.

Photo from http://webee.technion.ac.il/people/yoav/research/ACam.html

There was also talk about new CMOS sensor, which does such things internally, so cool,e.g. live object shape acquisition.

5. Ramin Zabih showed that “panorama stitching” task is far from solved and presented two methods, how to deal objects and moving things. Specifically — detect them :) Another interesting thing, that according to him currect state-of-the-art is Photoshop CS18 and difference to CS14 is huge.
And my lesson from his talk — don`t stick to common task definition, look at the real-world use-cases and what users do.

Examples of panorama problems because of parallax and/or moving objects and OOIS poster

6. Steve Seitz made a cool talk about Do it yourself 3D display from iPad + plastic. That video is not from workshop, but the technology is the same. Quite amazing :)

7. CNNs for aerodynamics from Pascal Fua was cool. Simple idea: train CNN to mimic simulator and/or real aerodynamic data. Now you have were fast and differentiable model, so can directly optimize shape of your wing/boat/bike. This was used for EPFL entry in some speed boat university competiton

Figure from “Geodesic Convolutional Shape Optimization”

8. Deep Internal learning from Michal Irani group. And Michal is just amazing. Who said that you cannot train on test set in test time?

Figure from https://arxiv.org/pdf/1712.06087.pdf

You can and this only test time training is faster (for high-res) and better for super-resolution and denoising, than GANs and other ResNets. Idea is to use fractal image self-similarity — old idea from her group.

There were more interesting talks, but let`s leave them for the next post :)

Computer Vision researcher and consultant. Co-founder of Ukrainian Research group “Szkocka”.