iiinova — where your digital self keeps evolving

Johan Edstedt @parskatt.bsky.social · 02/09/2026

Could have a lunch, not sure im a friend but I do exist.

Yep. There are also many implementation details that matter quite a bit. Some RANSACs e.g. do a final bundle while others don't. This matters a lot for benchmarks (but likely not at all when part of a larger SfM framework).

010

Shameless plug for my thing I did a while back

210

As pritt pointed out. Given correct hparams most perform similarly (if avg over diverse data)

Very off that gc ransac would be 10 pts above other ransacs. As usual hparam smell.

210

Not sure where I'm going with this, but I feel like this hints at a larger pattern of issues regarding how warnings are communicated to the public.

030

I'm somewhat concerned that this type of messaging would drive people to avoid the monitored lakes and assume the unmonitored are fine (not unreasonable, as finding the actual numbers required digging through pdfs).

There are multiple other nearby lakes that, IMO, seem much more likely to be heavily polluted.

100

It turns out that the lakes with warnings are basically the only ones that are being actively monitored.

Recently read an environmental warning (PFAS mainly) about eating fish caught from some nearby lakes. I was curious so I checked the source:

Johan Edstedt @parskatt.bsky.social · 20/07/2026

Constant jac det is a pretty strong constraint, and holds for degree 2. Not at all an expert but I guess the intuition is from linear maps which have constant jac det, and hoping it generalises.

Johan Edstedt @parskatt.bsky.social · 20/07/2026

Interesting that it sends 3 pts to same, would have thought it would be easier to find 2 pts to same. Will all counterexamples be that form?

Johan Edstedt @parskatt.bsky.social · 24/05/2026

So you support datacenters consuming all our oceans???

120

Johan Edstedt @parskatt.bsky.social · 19/05/2026

Functional programming be like:

050

Johan Edstedt @parskatt.bsky.social · 14/05/2026

deal with it surely should include sunglasses?

Johan Edstedt @parskatt.bsky.social · 08/05/2026

If its not backwards breaking its not worth doing.

010

Johan Edstedt @parskatt.bsky.social · 06/05/2026

Idk, my reviews for bullshit papers are very brief. Just clearly state main issue and recommend reject?

020

Johan Edstedt @parskatt.bsky.social · 23/04/2026

Yes, but it doesn't matter. Note: this comes from DKM lol.

020

Johan Edstedt @parskatt.bsky.social · 17/04/2026

Can't take Matt Damon + Tom Holland seriously.

Johan Edstedt @parskatt.bsky.social · 15/04/2026

Wait should have went with twoview geometry

000

Can expand further if you're interested, also @bokmangeorg.bsky.social is likely less inebriated and likes to talk equi.

020

I'm currently a bit drunk but I think what you're saying is basically ours if applied on only one image instead of pairs? The descriptor loss is on softmax(z, z'). If z' is just rot(x) then it's ours minus perspective? The matcher loss is similar but actually per layer, so even more sim.

Accepted to the Image Matching Workshop at #CVPR26!

1181

On HardMatch we particularly improve on the celestial subset, enabling new usecases for matching.

Qualitatively and quantitatively our features and matcher shows strong invariance properties.

230

Surprisingly, we found that invariance at the descriptor level is as good on upright, and enables smaller matchers!

We investigate whether rotation augmentation alone is enough for invariance, and at what stage to incorporate it (early or late)

140