In real-world robotics or autonomous vehicle use, vibrations, shocks, temperature changes, etc. can misalign the relative position of stereo cameras, so I'm looking into online stereo extrinsics self-calibration. Via Blender and Python I can get pixel-perfect keypoint matches.
Someone turned a bunch of iconic movie scenes into Gaussian splats. I've been thinking for a while which iconic movie shots would lend themselves best to this, i.e. long and stable tracking shots through scenes with little to no movement.
Went hiking today with a lovely group of friends, so we decided to take a group 3D Gaussian Splat instead of a group photo. Nice job staying mostly still. ;)
@LumaLabsAI
Playing with self-attention in latent diffusion models. This animation illustrates that the model learns to represent 3d scene properties like depth and object semantics. The red dot shows which pixel's self-attention map we're seeing.
#Diffusion
#DeepLearning
#StableDiffusion
Yesterday I got to present the
#GaussianSplatting
paper by
@Snosixtytwo
,
@GKopanas
and their co-authors to our seminar on Advanced Topics in Vision.
The dandelion on made for a lovely title slide.
I'm learning about normalising flows and flow matching and found these fantastic animations on . Not sure who created them, possibly the FFJORD authors ()? Great visualisations!
"mrcal is a generic toolkit built to solve the calibration and SFM-like problems we encounter at NASA/JPL"
Great documentation, tutorial, and tools for camera calibration:
My friend Daniel took these photos over Zurich last night while I was sleeping. Daniel, please burst in my door with an axe and drag me outside next time this happens! 🥹 I thought this stuff only happens to fantasy elves and Norwegians.
"A.I.-Generated Garbage Is Polluting Our Culture"
Will we act against the pollution of our information environment as halfheartedly and ineffectively as we've acted against the pollution of our natural environment?
On the left we see the disparity maps from the initial calibration. However, the stereo rig has since been de-calibrated through shocks and vibration during use. When trying to perform per-frame camera extrinsics re-estimation, I get more complete per-frame disparity maps but
Went for a hike yesterday and came across this pretty little Tell's Chapel by the lakeside. Chucked five photos into
#DUSt3R
. Looks like the benches would require more.
Exchange with my girlfriend:
"I hate it when DL-based methods that are essentially magic work better than well-understood algorithms in 3D vision."
- "Well then you're likely going to hate the next decades of your life." 🫠
Awesome looking course by
@vincesitzmann
. Thank you for making this publicly available! :)
(Don't get discouraged by the "broken" first recording, the subsequent ones work fine)
@yuliangxiu
@ETH
The rumours around this tragedy are deeply unsettling and I hope the circumstances of her passing are investigated with great urgency and diligence.
The chatGPT Mac app is pretty neat. It needs an option to increase the font size, but being able to start a prompt just like a Spotlight search (via ⌥ + SPACE) is very convenient!
@xchester16
@LumaLabsAI
If you're interested in the research, the techniques are referred to as Neural Radiance Fields (NeRFs, ) and 3D Gaussian Splatting ().
In practice you can use the Luma app to create 3D Gaussian Splats:
"A.I.-Generated Garbage Is Polluting Our Culture"
Will we act against the pollution of our information environment as halfheartedly and ineffectively as we've acted against the pollution of our natural environment?
@RadianceFields
@Snosixtytwo
@GKopanas
It wasn't recorded yesterday but I might record it myself and put it online some time in the coming weeks.
It's only a minor additional effort for me but might serve many more people than just the couple dozen that took the in-person seminar.
@CSProfKGD
It's funny – I wrote my BSc thesis in theoretical CS, and undergrad algorithms courses are of course all about discrete maths and graph theory. Now in my master's in Visual Computing and ML I use almost none of that. Maybe physics would've been a better undergrad degree for the
@cmuratori
I say that as a current CS MSc student but the problem is that knowing somebody has a CS degree gives you very little information. You can meet three people with CS degrees and it could be that person 1 had lots of coding and software application engineering courses but virtually
This looks like a wonderful complement to OpenAI’s “Spinning Up” tutorial introduction to deep reinforcement learning, with a somewhat different but equally interesting structure.
Despite its relevance, it can be hard to get into RL. This motivated us to write "An Invitation to Deep Reinforcement Learning", a tutorial for readers with only basic ML knowledge. Huge kudos to
@bern_jaeger
for this effort. Help us to spread the word!
Academia: “We propose the first Model that predicts the 3D model of an object from a single input image within just 1 second*.”
…
*) when using five NVIDIA A100 GPUs costing $50,000
#NeRF
Synthetic data allows us to get highly realistic RGB, and perfect depth, normals, motion vectors, and any other geometric ground truth data we may want for 3D (or 2D) computer vision. I'd love to see the CV and CG communities combine into "Visual Computing" (my MSc major). 👁️❤️👾
@rms80
@DummblondGaming
Agreed. Feels like it says "Let's crush your creativity by destroying all the tactile and haptic richness of the art utensils and instruments here, and replace them with a slab of glass." I'm very fond of my iPad but this ad missed the mark. Creatives like physical objects.
I'm reading a dissertation where virtually every single sentence is written in the passive voice, and sentences are way too long. Telltale signs that the author is a native German speaker.
My gripe with AI is that it won't tell me if the premise of a question is flawed. Asking it to explain parts of a mathsy paper/lecture often works. But if there's a mistake in the material, it tends to make up some BS explanation instead of saying "This is wrong and here is why".
@jon_barron
@giffmana
@ylecun
@geoffreyhinton
I believe I agree more with Yann than with Geoff but most of all I think it’s just really unfortunate and counterproductive that “AI risk” is now increasingly synonymous with “extinction risk”. There are huge risks I see with AI (and huge benefits) that are not about extinction.
@LMSacasas
I partially agree. The internet does deplete our will power and, via the many ways it shapes our culture, fragments our time and attention so that it becomes way harder to do anything that would require more activation energy, time, and attention, than mindlessly browsing the
I knew they should've gone with
⚠️ "Scientists SHOCKED: Large Language Models Now Fit in Your Pocket!" or
⚠️ "Memory Companies HATE Them: Researchers Discover One Weird Trick to Shrink LLMs" or
⚠️ "You Won't BELIEVE What These AI Researchers Did with Limited Memory!" instead.
Academic texts can be exhausting for opposite reasons:
(a) Each paper needs to repeat a lot of blah boilerplate.
(b) Papers take things for granted that you don't know.
Imagine an AI that remembers what you know, and that curates and helps you understand what you don't yet know.
@random_walker
2022: AGI will be a deity that shall conquer the galaxy.
2023: AGI will replace all humans at everything.
2024: AGI is a program that's useful for many things.
Marigold has such a pretty colourmap for its depth maps. I much prefer the red-to-blue gradient over the more commonly used white-to-black gradient for depth.
Hidden state in Jupyter notebooks is terrible. It's crazy that we work like this. I wish reactive Pluto notebooks were the norm. I want to be able to run any cell an arbitrary number of times, in any order, and know exactly what result I'll get.
@blazer_adam
We'll see.
@KBingxin
,
@AntonObukhov1
et al. just showed how to get high-fidelity depth maps out of SD. For relighting we need surface normals and it's still unclear whether SD learns those. suggests that SD does encode "same plane" relationships, so maybe.
US politics is so utterly confused. I’m not a fan of TikTok but this is such embarrassing protectionism, lobbied for by Meta. Pass a universal data privacy law that affects US social media companies equally, or get off your high horse.
@jon_barron
Yeah, control is the big one here. Waiting (or working) many hours for seconds of final footage is standard procedure even in live action films, not just Pixar animation. People have no idea how long it takes to shoot a scene that ends up being 60 seconds in the final film/show.
One nice thing about studying at ETH is when physicists who also happen to be world-class triathletes casually tell you that “a non-productive day is not the end of the world.” 🫠
@CSProfKGD
This was part of my initial exploration for our deep learning course project this semester. I’ll have a project report by the end of the semester. Just thought I’d share this because I found it cool. I’m new to DL (and research projects), so don’t know if it had been done before.
Fantastic TED talk by
@profgalloway
. The median age in the US is ~38.5 years. I wish someone would give this talk in Germany where the median age is 44.9 years.
I wonder if universities could create a custom GPT for each course to provide students with a tutor. Supply the GPT with all course materials, related readings, problem sets, and instructions, and let students use it as a 24/7 available TA.
#education
#chatGPT
#GPTs
#openAI
If Twitter could just allow me to auto-block any user (bot) profile that has a link to a camgirl site in its bio, that would go a long way. I just purged my follower list of ~40 obvious bot accounts and I'm sure I've missed many more.
Playing with self-attention in latent diffusion models. This animation illustrates that the model learns to represent 3d scene properties like depth and object semantics. The red dot shows which pixel's self-attention map we're seeing.
#Diffusion
#DeepLearning
#StableDiffusion
"The proton is a quantum mechanical object that exists as a haze of probabilities until an experiment forces it to take a concrete form. And its forms differ drastically depending on how researchers set up their experiment."
Sophon lock confirmed.
@yongyuanxi
@ZGojcic
The more time I spend among academic researchers, the more I understand why the German word for “curiosity” is “Neugier”, as in “greed for novelty” (neu = new, Gier = greed). I also notice it in myself.
I like that you point out the tension between self-serving and
I created this graphic to give myself an easily glanceable overview of the U-Net in the Stable Diffusion architecture. Maybe it can be useful to others as well.
#StableDiffusion
#diffusion
#LDM
@janusch_patas
When you first install it, you're shocked how terrible it is.
Give it a couple of days of active use and you'll be shocked by how well it manages to keep you hooked.
I had to uninstall it after a few days for that reason.
@JustinMSolomon
@sp_monte_carlo
I believe it's just a remnant of the standard DL introduction where you go from a Perceptron to an MLP, and these computational graph diagrams then justify the term neural "network." I also really don't think that narrative is worth it though.
@nathanbenaich
@RikeFranke
@ylecun
The saying "He has forgotten more about X than I'll ever know." means that a person is such an expert that even the volume of X that he has forgotten exceed the speaker's current knowledge of X. It's not diminishing LeCun's expertise, it's complimenting it.
Cool example of neural rendering using a model trained specifically for this game. Inputs are ball, paddle, and light positions (7 floats), output is the per-object texture (256x256). Training data created with Blender cycles.
@YiMaTweets
@young_opsimath
I agree in principle but in practice unfortunately not all lines of inquiry lend themselves to rigorous experiments. And logical deduction reaches its limits pretty quickly as the complexity and stochasticity of the system under observation increases.
It doesn't help that the main business pitch for AI/ML is that it can/will automate away much of what people currently earn their living with. Maybe once enough people flee into ML, the AI crowd will learn what it means to have their market destroyed by cheap and abundant labour.
Most people nowadays are interested in AI/ML because of the job prospective, or because it may lead to the growth of capital.
These are people who either hope to learn about ML so that ...
(2/10)