Dominik Engel

PhD in Deep Learning fron Ulm University. Interests: Deep Learning, Computer Graphics, Computer Vision. Account exists for DL research related Tweets only :P

[Leon Sick @leons1ck]

Accepted at #ICCV2025! 🎉 Many thanks to my amazing co-authors @xeTaiz, @Kopetri, @Pedro_He_Ca & Timo. Looking forward to the conference in Hawaii!🌺

[Pedro Hermosilla @Pedro_He_Ca]

Are you tired of projecting 2D features into 3D for general problem solving? Try our self-supervised model, which generates semantic off-the-shelf features natively in 3D! #CVPR2025

Project: https://t.co/jI8ZcpuXqY
Arxiv: https://t.co/HyjkrCpjSO
Github: https://t.co/d1qh5GLsa8

[Andrea Tagliasacchi @CVPR @taiyasaki]

📢📢📢 "𝐑𝐚𝐝𝐢𝐚𝐧𝐭 𝐅𝐨𝐚𝐦: Real-Time Differentiable Ray Tracing", a mesh-based 3D represention.

https://t.co/wcI6Xj6UHR
https://t.co/JOzRTfkgbl

Co-lead by my PhD students Shrisudhan Govindarajan and Daniel Rebain, and w/ @kwangmoo_yi

[Leon Sick @leons1ck]

🚨Super exited to share our new work CutS3D for unsupervised instance segmentation!
👀TLDR We connect semantics and 3D to cut instances and train a detector with spatial confidence.

🤗 Paper Page: https://t.co/d64oM1gaYj
🤗 Space: https://t.co/Ug7qacOBiZ
1/5

[Dominik Engel @xeTaiz]

Great work and charming presenter confirmed ;) come check it out!

[Leon Sick @leons1ck]

Can't wait to present our research on depth-guided unsupervised segmentation at #CVPR2024 tomorrow! 🙌

Make sure to stop by our poster #336 tomorrow during the Wednesday AM poster session 👀

For more, check out our project page: https://t.co/MsqEgattDr

[Leon Sick @leons1ck]

Can't wait to present our research on depth-guided unsupervised segmentation at #CVPR2024 tomorrow! 🙌

Make sure to stop by our poster #336 tomorrow during the Wednesday AM poster session 👀

For more, check out our project page: https://t.co/MsqEgattDr

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

Memory Mosaics

abs: https://t.co/Gyf3x02AMN

Meta FAIR presents "a new learning system architecture, Memory Mosaics, in which multiple associative memories work in concert to carry out a prediction task of interest."

"memory mosaics perform as well or better than transformers on medium-scale language modeling tasks" (GPT-2-small size on a TinyStories-like dataset)

[Alex @a13xba]

endoftext is live on @ProductHunt

@a_a_cabrera and I have been building for the past ~2 months and would love to read your comments.

https://t.co/aaNPLL3Olo

[Sebastian Hartwig @Kopetri]

We are happy to announce our survey and taxonomy on text to image synthesis evaluation!
There are interesting differences between SOTA metrics and their capabilities of measuring text-conditioned image quality.
Check it out at: https://t.co/I02WM3h4zn

[Itamar Zimerman @ItamarZimerman]

New!🚨📰
Mamba is a cool, efficient, and effective DL architecture, but what do we know about Mamba? How does it capture interactions between tokens? Can it be the attention-killer? In our work, "The Hidden Attention of Mamba Models" we provide answers to these questions! [1/4]

[Sebastian Raschka @rasbt]

Since Mixture of Expert (MoE) LLMs are all the rage as of this weekend, thanks to the Mixtral-8x-7B release, here's a quick explainer. The figure below shows the architecture behind the Switch Transformer (https://t.co/6jowgQx0DV), a great intro to MoEs.

The model depicted in this figure uses 1 expert per token with 4 experts in total. Mixtral-8x-7B, on the other hand, consists of 8 experts and uses 2 experts per token.

Why MoEs? Combined, the 8 experts in a 7B model like Mixtral are still ~56B parameters. (Actually, it's less than 56B, because the MoE approach is only applied to the MoE layers, not the self-attention weight matrices. So, it's likely closer to 40-50B parameters.)

However, since the router reroutes the tokens such that only 7B parameters (instead of all 56B) are used at a time for the forward pass, the training (and especially inference) will be much faster compared to the traditional non-MoE approach.

If you read my AI and Open Source in 2023 article (https://t.co/C8SGseRHNZ) approx. 2 months ago, I mentioned that "It will be interesting to see if MoE approaches can lift open-source models to new heights in 2024". It looks like Mixtral started this trend early, and I am sure that this is just the beginning :).

[Leon Sick @leons1ck]

Happy to share our new pre-print on guiding unsupervised semantic segmentation with depth maps!📄👀

TLDR; We guide the feature space for unsupervised segmentation with information from the 3D space through depth-feature correlation.
HF Paper Page: https://t.co/dyuz0qXlCg
More 👇

[Dmytro Mishkin 🇺🇦 @ducha_aiki]

VisProg
#CVPR2023 Best Paper Award

https://t.co/9HN2yQlHs6

[AI at Meta @AIatMeta]

📝 New research from Meta AI — Hiera is an extremely simple hierarchical vision transformer that's both more accurate than previous models + significantly faster at inference and during training.

Paper ➡️ https://t.co/mPN3nPra6g

[EuroVis @EuroVisConf]

📜 Explore your large image datasets using their semantic content.
✍️ @a13xba, C. van Onzenoodt, @ResearchDaniel, and T. Ropinski
👉 Demo: https://t.co/v39eANlt9E
#Shortpaper #EuroVis #Eurovis2023

[John Nay @johnjnay]

Detailed LLM Evals

-Stratified eval can reveal subfields where hallucinations are more likely to occur
-LLMMaps: new visualization transforms Q&A data + LLM responses into internal knowledge structures
-Compares BLOOM, GPT-3, ChatGPT, LLaMa-13B, etc.

https://t.co/VMXGJTmTUZ

[AK @_akhaliq]

LLMMaps -- A Visual Metaphor for Stratified Evaluation of Large Language Models

abs: https://t.co/nryZiABEaM

[Dominik Engel @xeTaiz]

@huggingface @Docker How can I reliably retrieve a docker image name from a HF space URL programmatically? Mostly it matches the URL path with dashes instead of slashes, but there are some spaces that have slightly different names, i.e. (https://t.co/TWlwy0TqoN) ends with "webu-185266f" not "webui"

[Dominik Engel @xeTaiz]

@osanseviero @huggingface @Docker Thanks! You guys are 🔥