Klaus H. Strobl

Sensor algorithms robotics @DLR_en

[Klaus H. Strobl @5trobl]

@KostasPenn Middle layers for VGGT and DA3! https://t.co/SeudkqmZFw

[MC Squared @mcsquared34]

It’s almost like earth was perfectly created for human beings, literally water falls from the sky and food grows on trees, but somehow we find ourselves trapped in a machine that requires credit scores and a 40 hour work week for existence.

[Klaus H. Strobl @5trobl]

@softwire @karpathy Indeed. Whenever I come across a compelling interpretation of a ML method, my confidence increases from the prior 2% to merely 30% at most. So I'm probably still wrong, alas!

[Klaus H. Strobl @5trobl]

@fdellaert btw some implementations already avoid this approximation (as shown in the appendix for oss) but this nice paper addresses the issue in detail

[Klaus H. Strobl @5trobl]

@fdellaert In a way this is a simple form of @ajd's active matching paradigm where after every single "iteration" the state must be invariably updated for consistent stats ;-)

[Klaus H. Strobl @5trobl]

@AjdDavison The video from a camera attached to your children’s football being kicked is next.

[Klaus H. Strobl @5trobl]

@MargaritaChli @AjdDavison @TheBMVA Is that Gerardo giving him the award? 😂

[Klaus H. Strobl @5trobl]

@JoanSol20 @AjdDavison I remember laughing at this picture on delayed init #spot on

[Klaus H. Strobl @5trobl]

@elonmusk https://t.co/ZzcllAb4Yg

[Klaus H. Strobl @5trobl]

@pesarlin @veichta @PhilippCSE @mapo1 It’s definitely the right call to release it last in your research process. Also because of the low res in the current dataset. But we have to think big!

[Klaus H. Strobl @5trobl]

@pesarlin @veichta @PhilippCSE @mapo1 True, makes sense. A bit of human intuition is still needed.

[Klaus H. Strobl @5trobl]

@pesarlin @veichta @PhilippCSE @mapo1 In addition, perspective fields are robust to its shift by design. You surely tried? Didn't you see any effect or was it problematic for the LM to quickly converge to unique up-vector and latitude fields?

Cheers!

[Klaus H. Strobl @5trobl]

@pesarlin @veichta @PhilippCSE @mapo1 Hi Paul-Edouard, congratulations once again on your latest work and this time also on your career move!

When reading GeoCalib I saw so may parallels in its gist with PixLoc. And again, you didn't go for the title "Good features are all you need to learn for calibration" ;-)

[Klaus H. Strobl @5trobl]

@pesarlin @veichta @PhilippCSE @mapo1 I think you should go one step further and release the principal point in the model-based LM iterations. It is critical for high precision in traditional checkerboard-based calibration because distortion always dictates its location, hence it should help you as well.

[Klaus H. Strobl @5trobl]

@altryne "Easy Move+Resize" does the job just fine, as in Linux.

[Klaus H. Strobl @5trobl]

@Xallt_ @ducha_aiki If it’s synthetic data, generate at higher res than necessary first.

[Klaus H. Strobl @5trobl]

@ylecun Scale invariance is overrated

[Klaus H. Strobl @5trobl]

@francoisfleuret typos: an infinitely, constrains it

[Klaus H. Strobl @5trobl]

@araffin2 @esa @astro_marcus Amazing first time! And some last-minute novel moves!
But: Marcus was clearly not paying attention was he?

[Klaus H. Strobl @5trobl]

@freebsdfrau Thx. That might be the reason for the remaining 20% gap but I’m afraid a LAN port was configured to 100mbps 🙈 and ssh packet retransmissions were there pervasive.