Srujan Deolasee

Two months ago, I vaguely posted a number: 0.9 FID, one-step, pixel space. Now it is 0.75, and can be even lower. Many wonder how. I thought it might end as a small FID prank: simple and deliberate. It started with one question: can FID be optimized directly, and what does it reveal? Introducing FD-loss.

@Luckyballa I see, that makes a lot of sense. I might have been biased due to working with less noisy sensor data thus not needing much post-processing before building the TSDF. Def agree that it’ll help shape completion/smoothness, and I’ve wanted to tackle that since a while now

In my experience, to utilize any multiview signal effectively (depth, color, camera params), assuming they are not 100% perfect, you want a differentiable method All observations have to agree on what the scene is, not just by averaging out, but by measuring multiview consistency With a differentiable method, you can spot and even correct outliers without hacks, use priors and regularization for shape completion and smoothness and a couple other things

I see that a lot of papers and startups still use TSDF fusion to reconstruct surfaces from depth maps Still relying on voxel grids, memory constrained and low quality I think I should solve it next weekend

@Luckyballa Fair. What applications are you looking at for them to be differentiable though? Uncertainty can be thought of as how may times the voxel has been observed from multiple views, and I think libraries like open3d can give you a mesh easily too

I agree, sparse variants are not as bad memory wise Classic variants are non differentiable, washing out the details, don’t hold any notion of uncertainty, and if you want a mesh, you have to do marching cubes or similar, further increasing the error and complicating online reconstruction

INSANE: An 8 minute compilation of flops, falls, and 50/50 calls, (that all coincidentally went OKC’s way) in round 1 of the NBA playoffs vs the Suns… Enjoy 😂😂 (Via, @Hero_OfThe_Day)

@taiyasaki Ohh I did not know that either. Coming from robotics, I think I had mostly seen it in context of robot planning

@sruj_d Well, voronoi diagrams pop up in information theory as you discuss optimal quantization. So it's not that surprising that they make their way into representing other forms of signals in discrete form.

📢📢📢introducing 𝐏𝐨𝐰𝐞𝐫 𝐅𝐨𝐚𝐦 A 3D representation that can be ray traced or rasterized in real time, with NO COMPROMISE in quality. - Project: powerfoam.github.io - arXiv: arxiv.org/abs/2604.24994 Rasterized at 3DGS-class FPS Ray traced at Radiant Foam speeds

@taiyasaki While I’m no 3dgs expert, I never expected to read voronoi in this context lol

The recipe is Voronoi at every scale: - Bounded *power* diagram → 3D geometry (cells with controllable extent) - 2D Voronoi on each cell → texture / displacement - Spherical Voronoi on each texture site → directional radiance Three Voronoi... one differentiable representation

I don't work on reliability & scaling at GitHub, but the people who do aren't bad at their jobs. They're dealing with unprecedented scale from agents. It's easy to shit on GitHub from the outside if you're not in charge of 30X-ing capacity within a few months. Have some grace.

@aelobdog Kasa chalu ahe? Long time no chat

@sruj_d "BUT WAIT !"

On some occasions it gives me 50+ weeks timeline😂😭

you can now control things with your brain. literally. we're building the most wearable BCI on the planet, with @sabicap, backed by @khoslaventures @accel @initialized & @kevinweil. we collected the world’s largest neural dataset and trained the most capable Brain Foundation Model. then we invented a new class of biosensors powered by custom ASICs. type without typing. click without clicking. a cap that lets your brain do the work. we’re sabi.

Learn more about what we have been building at @genrobotics_ai 👇

Scaling robotics isn’t just better models or better infra, it’s making intelligence composable and ofc continuously improving.

the feeling of a successful hatewatch, but you’re lowkey next

@comma_ai Not compatible with Chevy Equinox 25

What’s holding you back from buying or recommending comma four?

It's 1:30 am and I've nothing better to do, so here we go: * stereo vs. mono: yes, if you have stereo, use stereo. If you're stuck with mono and have some compute to spare, use an ML depth model to get approximate depth. Accuracy at this stage is overrated. I'll all get bundle-adjusted anyway. * feature tracking: search vs. KLT. Both have upsides and downsides. The best systems use a hybrid. Search is often too expensive and depends on detection, which is often brittle -- but can save your butt in many edge cases. KLT is fast & robust but bare KLT is not very accurate and also drifts over time. I hope to be doing one or more posts on this very topic hopefully soon. * pose estimation: directionally correct but there's a world of best practices to make this fast & robust. People have written entire PhD theses about this. Topic for another post. * KF-based map expansion: yes that's best practice. But KF-selection based on "every few meters" is instant game over. Lots of cases and edge cases that need to go into a suitable heuristic. * CUDA kernels for stereo matching: serious overkill. Matching few hundred features on CPU takes at most a couple milliseconds if implemented right. * Local BA: 12 KFs is kinda arbitrary. Might work well for KITTY but not generalize. * Eval on KITTY: that's easy-tier: camera always upright. No pure rotations. Very controlled motion. Very large field of view. Drone datasets are where the rubber meets the road. * Performance: 9 FPS on RTX 3050. NGL, that is brutally, ludicrously slow. Us old-schoolers did realtime visual SLAM 20 years ago on ~ 1/1000th of the compute budget.

@Parskatt Played around with it quite a bit and I’m very impressed looking at performance gains! Congrats!

Introducing LoMa, the next generation of feature matcher!

Socha nahi tha samay ka video dekhke rona ayega. Can’t wait for latent to return

how were people using chrome till now? had no idea it didn’t have this already

Real