WildGaussians: 3D Gaussian Splatting in the Wild Jonas Kulhanek Songyou Peng Zuzana Kukelova, Marc Pollefeys Torsten Sattler tl;dr: splats with embeddings to get view-dependent color/illumination. Trained with DINOv2 features loss arxiv.org/abs/2407.08447

Still not sure why the ML community adopted conda instead of plain old virtualenv

TimDarcet I am using virtualenv. I have no issue installing different python versions or cuda. I hate conda deeply. I hate conda as much as jupyter notebooks.

Okay this uiua thing is actually pretty fun

Okay first DFN then that Apple is now the king of open-source datasets, both vision and NLP

Communication : Developed NCCLX, a fork of Nvidia's NCCL library, to improve collective communication performance, especially for higher latency networks and better debugability.

updated the website

Lmao they waited for the 405B release just to be able to 1-up it

About Self-Supervised Learning of image representations: - generative architectures trained by reconstruction sux. - Joint Embedding Architectures trained with various criteria r00lz.

Easy way to fix academia: Add a negative offset to all citation counts in Google Scholar. Every new paper starts at -100 citations and has to claw its way out of that hole.

(1/8) 🎬 Introducing the Short Film Dataset (SFD), a long video QA benchmark with 1k short films and 5k questions. Why another videoQA dataset? 📖 Story-level QAs 🎥 Publicly available videos 🔒 Minimal data leakage ⏳ Long temporal context questions shortfilmdataset.github.io

lol apparently X quietly and automatically opts everyone into being used as training data for Grok if you wish to opt out, here’s the link: x.com/settings/grok_…

Hear me out: a large hall with plenty of reasonably *comfy armchairs* and (hear this) *whiteboards*.

Theia: A Robot Vision Foundation Model that Simultaneously Distills Off-the-Shelf VFMs such as CLIP, DINOv2, and ViT itinai.com/theia-a-robot-… #TheiaRobot #VisionFoundation #AIforBusiness #EfficientTraining #AutomatingOperations #ai #news #llm #ml #research #ainews #innovation …

Wait wait guys is this... ???

actually this can be attacked from two sides: (a) require publicly posted AI to serve metadata, (b) NIST standard for digitally signed authentic photos

The mourning will be long, but the special status of "recorded signals" is no more. Repeat after me: images, sound recordings, and videos have the exact same validity as gossip as carriers of truth.

Nice! But also, a massive opportunity miss to remove all near-dups with common vision datasets like imagenet and mscoco to rid all future papers from that confounder once and for all 😢

>2004 >play wow and deal with OOM (out of mana) >2024 >play cloud infra and deal with OOM (out of memory)