Watching the RAG crowd getting hyped about old stuff for the past year has been fun 🙂 First, they were PUMPED about vector search. Startups were built just for that. Now, they discovered that BM25 can be better than embeddings. Waiting for them to discover Learning to Rank 😆

Re-LAION-5B fixes the issues reported by Stanford Internet Observatory (SIO) in December 2023 for original LAION-5B. In cooperation with IWF, C3P and David Thiel (SIO), all 1008 links to suspected CSAM in the report are removed from LAION-5B metadata, using safe hash lists.

Pretty good research paper.

Augmented Reality without Borders: Achieving Precise Localization Without Maps Albert Gassol Puigjaner, Irvin Aloise, Patrik Schmuck tl;dr: switch query vs db images - register db to query seq. "VisLoc from Ess.Matrices"+bells&whistles, but not cited? arxiv.org/abs/2408.17373

GMM-IKRS: Gaussian Mixture Models for Interpretable Keypoint Refinement and Scoring Emanuele Santellani, Martin Zach, Christian Sormann, Mattia Rossi, Andreas Kuhn, Friedrich Fraundorfer tl;dr: homography adaptation is not enough. Helps all on IMC21 arxiv.org/abs/2408.17149

HPatches users, do you know that there is a paper from the author of HPatches, a protocol and a repo for evaluating local feature DETECTORS on HPatches? Not those "matching scores vs LoFTR"? W/o googling.

Is there any kind of “dinov2-tiny”?

Reconstructing scenes and point tracking are two sides of the same coin. Existing methods require offline reconstruction or multi-view camera setups - unrealistic setups for real-world applications. Enter DynOMo! arxiv.org/pdf/2409.02104 🧵

Kaggle launched a really interesting tabular competition today that has some rich text features as well: Predict which variants of Monte-Carlo Tree Search will perform well or poorly against each other in hundreds of board games kaggle.com/competitions/u…

A recent breakthrough (and really beautiful maths) behind optimal transport and the Sinkhorn algorithm, used in many recent matching papers e.g. SuperGlue Paul-Edouard Sarlin Not really useful for ML but really fascinating imho youtu.be/-uIwboK4nwE?si…

UWStereo: A Large Synthetic Dataset for Underwater Stereo Matching Qingxuan Lv, Junyu Dong, Yuezun Li, Sheng Chen, Hui Yu, Shu Zhang, Wenhan Wang tl;dr: in title. arxiv.org/abs/2409.01782

RoomDiffusion: A Specialized Diffusion Model in the Interior Design Industry Zhaowei Wang, Ying Hao, Hao Wei, Qing Xiao, Lulu Chen, Yulong Li, Yue Yang, Tianyi Li tl;dr:a story of dataset curation and model training for domain specific diffusion training arxiv.org/pdf/2409.03198

600 influencers is a crazy number. A few hundreds of millions of dollars will buy you an entire social media influencer ecosystem -- apparently this must be very good ROI for the Kremlin given its willingness to pay millions to B-listers (who'd sell out for a fraction of that)

I second that. Before we introduce draconian measures, we should get a sense of the scale of the problem. My impression is that unresponsive reviewers are by far the most prominent problem. Followed by unhelpful hand-written reviews. LLM reviews somewhere down the list.