okay, playing around with your suggested packages takes a lot of time... aand I'm not sure if this is time well spent.

but it looks cool :)

I think It's possible with manual optimization, but I agree with you that for most non-standard deep learning pipelines lightning does not work well out of the box and can feel limiting.

lightning.ai/docs/pytorch...

opinion on lightning?

I suppose most people are aware, but for most only this personal exposure to directly see what your research is used for (your experience with Samsung) will you to actually consider if you want to continue working in the field or not. (Also depends on your own moral compass of course).

I never really used it, but given that RANSAC is designed to handle "data containing a significant percentage of gross errors", wouldn't using sigma take away the original usecase of RANSAC since it is affected by the outliers? I guess it depends on the application and how much outliers we expect...

How does it lead to threshold-free RANSAC? Fit model X times and choose the one with lowest inlier noise (then we still need a threshold to choose in/outlier)? Maybe I (also) don't know enough about RANSAC to understand it :D

also lots of political content landing in my feed unfortunately...

Which one is your favorite?

How long does a PhD usually take in Sweden? I assume you need a Master's to apply, so the system is similar to Germany and many other European countries (BSc (3 years) -> MSc (2 years) -> PhD (X years))?

I'm not saying you have to agree with his conclusion, but it at least made me think of where we might be going in the future.

And Would Transformers be any better here?

But doesn’t that mean that CNN-based monodepth predictors are kind of useless for any real-world application where we are out of distribution (either in terms of resolution, or absolute depth eg indoor vs outdoor scenes?). Sure, we can pre/postproces, but does that fix it?

okay, that was a stupid question, the answer was right in the comments 🫣… But it makes sense, thx!

But still, we do we observe such different results?

Awesome, I’m looking forward to it :)

Sounds very exciting. How flexible is the timing? Both duration and start/end dates

what makes you feel that way for Computer Vision?