Titus Cieslewski

@maththrills Huh. Related to what @AlexBewleyAI points out, maybe this is related to why predators have restricted FoV VS prey?

@hermannsblum @_krishna_murthy @Publons Interesting, good to know, thanks! Sounds easier to implement, but indeed not as easy to adopt...

@tcies @_krishna_murthy @Publons The same is possible with ORCID as a more open system. Unfortunately, ORCID requires the journal/conference to update the reviewer‘s record directly and there is not much adoption yet.

I think @Publons is a criminally underrated / underutilized tool. It collects verified reviews into a public record. If more people used it, we could have a better conversation about review duty. Thoughts?

@ducha_aiki @_krishna_murthy @Publons Interesting, thanks! A priori it seems to me like w/o alternatives advantages outweigh disadvantages (& they need some business model). Peer review anonymity is not 100% watertight anyway. Worried about "malicious" leak or f*ck-up? What's wrong with publishers motivating you?

@tcies @_krishna_murthy @Publons I always decline Publons request, when do a review for the IJCV. The reason: I don't trust it for not opening my identity as a reviewer. Plus: I don't want publishers to motivate me.

@pesarlin @Publons I would guess so. So far I have not encountered a review that wasn't approved. It's super simple, you just need to forward a review confirmation e-mail (make sure one is sent, in CMT this is an option afaik) to an e-mail address they provide.

@tcies @Publons Does it work with any journal/conference? Or does it need to be pre-approved?

Very nice interview with @SattlerTorsten about Computer Vision between Geometry & Learning, and research trends. Personal takeaways from first ½: AR/VR as stepping stone for robotics; visualization is key to understanding; simulations allow agents to make and test hypotheses.

@todoist Awesome, thanks a lot for your amazing work 😊!

@tcies Hi Titus! Thank you for sharing your feedback! This has been noted📝

Hey @todoist , it would be really nice to have "negative priority" / p5 and maybe p6. So that in my "Today" view, I can move "less important than default" tasks back without having to mark all default tasks as higher priority. Should be trivial to implement.

@fdellaert Yes! Recently switched from paper to digital annotations inspired by @wuningxi 's slides and this is so tedious! Though I think some kind of built-in split-screen would be even more important for this particular problem. Please @paperpile 😬?

I find myself constantly flipping to the references page.

How about giving all conference submissions one extra page but insisting we use verbose citations, like (Mildenhall et al. 2020).

@deviparikh @CSProfKGD Is that a problem 🙂? "The fifth element" totally called this one: youtube.com/watch?v=ViCNgK…

Anyone else worried that after watching so many recorded talks at 2x the speed, you'll be impatient when we go back to watching talks in real-time?

A candidate for "things that will be different after Covid"? Well, why not?

@ankurhandos I saw this a couple of weeks ago in this neat YouTube video: youtube.com/watch?v=thOifu…

Maths is a universal language.

Time and attention are precious, and we should treat each others’ time and attention with respect. Here is a proposal for some 21st-century rules that should help us do that with modern communication methods: titus-c.ch/2020/12/respec… Thanks Carmen and @foehnph for reviewing!

@CsabaSzepesvari I'm not at NeurIPS, but this sounds like a fun programming / automation challenge 🤔.

NeurIPS experience: Does anyone enjoy moving around a silly avatar with the speed of a snail in oversized rooms to get to specific posters?

@mikko_lauri Thanks! Reliable and fully connected (e.g. multi-hop, see work by @af_robotyk for connectivity. maint.). Single failure can be handled w/ redundancy @ extra cost. Temp. failure should be ok. Assumption only affects work on O(n) dec. place rec., irrelevant for other contribs.

Ph.D. Defense: Multi-Agent Visual SLAM with minimal data exchange: a) O(n²) → O(n) decentr. place recognition, b) minimal feature set with SIPs, c) no more feature descriptors with IMPs, d) optimization is not needed for navigation, nor exploration: youtu.be/5C7I7uNGkBw

@ducha_aiki @jatentaki Good idea! When using cross-entropy, the detector score can even make halfways decent predictions about whether detected points will be correct matches 😉: arxiv.org/pdf/1805.01358…

Hi, @jatentaki, Q about DISK and RL. Imagine I would like to train detector by prediction if the keypoint is a correct match and use CE. And to train descriptor the same objective based on descriptor distance. Wouldn't the loss be the same, as in DISK but without mentioning RL?

@fdellaert As an aside, I remember that you once said, when visiting ETH, that the contributions that stick around are the simple ones that everybody can understand. This inspired a lot of my "problem picking" choices!

@tcies Nice! Congrats! Paper or thesis link?

@fdellaert Thank you, Prof. Dellaert! I will upload the thesis after addressing the reviewers' comments. The core papers are linked in the YouTube description: youtu.be/5C7I7uNGkBw

@DesaiDjd Thanks! I will upload the thesis after addressing the reviewers' comments! In the meantime, a lot is already in the core papers (see YouTube description).

@tcies Congratulations, if possible can we have thesis link it is interesting to read