Eric Brachmann

📢A new learning-based approach to SfM: #ACEZero No img-to-img matching, optimises image-to-scene correspondences directly. Needs no pose priors. Works on unordered image sets. Efficiently handles thousands of images. Paper: arxiv.org/abs/2404.14351 Page: nianticlabs.github.io/acezero

@chrisoffner3d @ducha_aiki Not sure about that, see my other answer. "More forgiving" maybe... Demands proper analysis... 😉

@eric_brachmann @ducha_aiki Great, thank you! Yeah, my intuition is that NVS is a lot more forgiving than SfM and 3D reconstruction.

Scene Coordinate Reconstruction: Posing of Image Collections via Incremental[...] @eric_brachmann Jamie Wynn @ShuaiC8 Tommaso Cavallari @AMonszpart @dantkz @viprad tl;dr: RGB+ZoeDepth for init->RANSAC+pnp->fit scene coord regression -> refine&repeat arxiv.org/abs/2404.14351 1/

@chrisoffner3d @ducha_aiki Pose benchmarking is a problem still looking for a good solution.

@chrisoffner3d @ducha_aiki For the record, I have seen something similar in visual relocalization benchmarks, where a relocalizer mimics the defects of the SfM pseudo ground truth to get deceivingly good results in a broken reconstruction. That's one reason I do not believe on COLMAP pseudo GT.

@ducha_aiki @janusch_patas Indeed. Although it should be coupled with a proper train/test split of disjunct camera trajectories. The usual 8th frame evaluation is far too easy and can be very deceiving. I have the impression even more so for GS than NeRFs.

@janusch_patas Tagging @eric_brachmann as he did exactly that, but with nerfs for AceZero)

Hot take: Pose estimation methods should generally be evaluated using a baseline training method, such as 3D Gaussian Splatting, to assess novel-view synthesis quality.

nianticlabs.com/careers/openin… Also apply if you fancy topics like depth estimation, meshing, splatting, vision-language models, HCI... Niantic Research works on a wide range of topics related to the future of AR.

Niantic is looking for research interns for next summer in London. Interested in the next generation of reconstruction, mapping and visual relocalization? Apply! (And send me a DM for good measure). Let's explore what the bitter-sweet lesson can do for 3D vision. Link below.

Time to move...

@CSProfKGD 55,935

Another 55,936 followers to go 🤓 Good to have options but not going anywhere.

After all these years, I still enjoy submitting to conferences. Is it polishing text late at night? Arguing with reviewers? Adding another publication to the pile? No. It's because each submission is a testimony, an 8-page certificate, that I learned something new. #CVPR2025

@Parskatt Within tight confines, Grace blooms where rules carve the way— Freedom forged by bounds. (haiku written by ChatGPT)

That papers should be *exactly* 8 pages is seriously stupid

We wrote a text on how we envision the future of visual mapping: nianticlabs.com/news/largegeos… So far, we have trained ~50 million ACE networks. 😎 Together, that's 150 trillion parameters... Not bad! But the future lies in going from local small models to a global large model.

@david_picard :D

@eric_brachmann Ah, the worst ones are the ones with the "GB" sign on their cars. The famous saxon "Griminal Bolizei".

Slow morning today. The compute cluster is in maintenance, time to go through my colleagues' comments on my little summer project (which is very much delayed).

@david_picard Do not underestimate the Stollen police of Saxony.

@eric_brachmann I don't care about the rules. I'm 1000 kilometers away anyway, so nobody will catch me.

@david_picard @ducha_aiki That was not the intention of my tweet (a desire for a meticulous references section was) - but reading my fair share of sloppy arXiv tech reports lately certainly reinforced my desire to have some sort of stamp of approval that verifies soundness before I invest reading time.

@ducha_aiki @eric_brachmann Because people need to know if it has been peer reviewed and accepted or not before leveraging it as a weapon in a review.

Dear author, thanks for providing a bibtex entry on your project website. How about updating it with whatever venue the work has been accepted at since the original arXiv report?

@jcivera @MattNiessner @LukasHollein From the Nerfbusters paper:

@jcivera @MattNiessner @LukasHollein Sounds like visual relocalisation datasets, that have designated mapping and query sets, would be a good choice when evaluating NVS. The Nerfbusters paper (ethanweber.me/nerfbusters/) was also making the same observation, and basically recorded a reloc dataset as a result.

NeRF and 3DGS reconstructions always look amazing in research papers but getting good results with existing 3D scanning apps is tricky. Why? Papers tend to show NVS results on test camera trajectories that are quite close to train views. Thanks @LukasHollein for the data :)

@lbertinetto A classic.

@eric_brachmann "Let's play a game we call _Name That Dataset!_" :)

As you write your #CVPR2025 papers, think of captivating first sentences. Example? "Creating noise from data is easy; creating data from noise is generative modeling." (from Song et al., ICLR21, arxiv.org/pdf/2011.13456)