Duncan Frost

Computer Vision Researcher at @RealityLabs. Previously at @Scape. Previous PhD student and Post-Doc at @UniofOxford

[AI at Meta @AIatMeta]

Sonata: Advancing Self-Supervised Learning for 3D Point Representations

Sonata represents a significant leap forward in the field of 3D self-supervised learning. By identifying and addressing the geometric shortcut and introducing a flexible, efficient framework, it’s an exceptionally robust 3D point representation.

This work advances the state of the art and sets the stage for future innovations in 3D perception and its applications.

Learn more ➡️ https://t.co/xazXDMdoHQ

[Xiaoyang Wu @XiaoyangWu_]

📢Sonata: Self-Supervised Learning of Reliable Point Representations📢

Meet Sonata, our"3D-DINO" pre-trained with Point Transformer V3, accepted at #CVPR2025!

🌍: https://t.co/jIte27gVmW
📦: https://t.co/qqEOChCZ86
🚀: https://t.co/GQ71nMwIVu

🔹Semantic-aware and spatial reasoning representations learned with no label;
🔹3x linear probing accuracy (from 21.8% to 72.5%) on ScanNet;
🔹2x data efficiency performance with only 1% of the data compared to previous approaches;
🔹As always, establish new SOTA results across indoor and outdoor 3D perception tasks.

Our author team: @HengshuangZhao, @jstraub6, @rapideRobot, @ddetone, @NinjaDuncan, @TianweiS, @Christopher_Xie, @NanYang719.

[Qiao Gu @qiaogu1997]

🚀EgoLifter is open-sourced: https://t.co/DjXB80aZel!

It includes:
🕶️Full recipe for Project Aria datasets, from downloading, pre-processing, to vignette handling
🌟Rewritten 3DGS train/test pipelines using lightning and gsplat
🌐Web-based 3DGS visualizer powered by viser

[Qiao Gu @qiaogu1997]

📹 On several large-scale egocentric video datasets, EgoLifter also showcases the ability to learn consistent 3D features and decompose a 3D scene into 3D object instances, which opens up promising directions for egocentric video understanding in AR/VR applications. 🕶️

[Edward Miller @TweetEdMiller]

📢 Today, our team is excited to announce 𝑺𝒄𝒆𝒏𝒆𝑺𝒄𝒓𝒊𝒑𝒕: an end-to-end method for scene reconstruction from @RealityLabs Research, which enables AI agents to understand and describe 3D spaces.

https://t.co/Te0pNtkFfS

/🧵

[Duncan Frost @NinjaDuncan]

Details and official job link can be found here: https://t.co/kZjM5uLb9v

#PhDInternship #MachineLearning #TechCareers #AI

[Duncan Frost @NinjaDuncan]

Are you interested in 3D scene understanding and reconstruction? Does the challenge of tackling these problems using egocentric image data sound exciting to you?

[Duncan Frost @NinjaDuncan]

Our team at Meta Reality Labs Research is looking for PhD candidates with experience in 3D computer vision and machine learning to join us for our Summer 2024 internship program.

[Matthias Niessner @MattNiessner]

Often AI methods justify their design by "That's how human's do it".

Not always a good idea: 1) we don't fully understand humans. 2) Think about how bad humans are doing basic arithmetic operations like addition or subtraction, and how well computers can do it in comparison.

[Matthias Niessner @MattNiessner]

The main idea of Machine Learning is that everything is simply learned from data instead of tedious handcrafting features.

The irony is that we are now doing more empirical research than ever, spending most of our time with data engineering and hyper parameter tuning.

[John Carmack @ID_AA_Carmack]

Docker containers are kind of neat. They are also kind of a craven surrender to the rotting mess of excessive software complexity.

[Davide Scaramuzza @davsca1]

Congratulations to NASA and all their collaborators! First powered autonomous flight on another planet successful! Btw, it is also a vision-based flight (single-camera + IMU + laser altimeter).

[NHS Million @NHSMillion]

NHS staff are being stretched as thin as they possibly can before they break. We had 38,000 vacancies for nurses and 10,000 for doctors even before the pandemic. We must look after our staff now or risk losing them when the dust settles. RT if you agree.

https://t.co/gyBdZu9TDL

[The Intercept @theintercept]

Argentina, South Africa, Brazil, and Turkey — all countries that hosted vaccine trials — will have to be satisfied with Pfizer’s gratitude, because (like most countries in the world) they won’t be receiving enough of the vaccine to inoculate their populations anytime soon.

[Yannic Kilcher 🇸🇨 @ykilcher]

GPT-3 is AGI

[Josie Zayner @josiezayner]

This is dum

At least an adult has a chance to overcome poverty and social class. A kid has zero chance. This is basically a "most privileged kid of the year"

I imagine if I was judged by my life at 15. Broken home, failing in school. Forbes 15 under 15 coming next 🤮

[Duncan Frost @NinjaDuncan]

@monge_maurizio you may find this interesting ;)

[Torsten Sattler @SattlerTorsten]

Very cool project presented today at #3DV2020: Kyle Wilson, Scott Wehrwein, Visualizing Spectral Bundle Adjustment Uncertainty, https://t.co/q1hTtte2ae
They have very interesting visualization https://t.co/kzEZEOPSvN
@fdellaert I think you would like this.

[John Delaney 🇺🇸 @JohnDelaney]

Pay everyone $1500 to get vaccine

It is a stimulus check & big vaccine incentive rolled into one

Result: more people get vaccine, it saves lives, helps people financially, and the $400 B pays for itself w/ economic boost from ending pandemic

RT if you agree #vaccinestimulus

[Hardip Barot @Hardip_nirvana]

So we’re slowly approaching tier 4 lockdown! By the time of Christmas 🎄 we’ll be in Tier 12 which means everyone will have to self isolate from within! You will not be allowed to look in the mirror or talk to yourself. Tier 12 is when you let your soul free.

[🌞🌵 cactus cowboy 🦀 @_cactuscowboy]

@seanmcarroll as an avid listener of the podcast, I was curious if you’re open to suggestions for future interviewees? There are a couple I’d really love to hear you speak with, in particular this mathematical physicist at Lund https://t.co/guZ5hE0ckF