Sometimes the best solution is a well-established geometric principle from decades ago. 🤝 📄 Paper: https://t.co/mZ5MAcx7Tf 💻 Code: https://t.co/jIwiv0zyrz Check out the project page for comparisons and results!

Training on static scenes with dynamic cameras surprisingly generalizes to dynamic scenes. This makes sense because stable camera trajectories reduce artifacts across ALL scene types, even with moving objects.

The key insight? Classical geometric constraints provide cleaner optimization signals than modern learned metrics. We found that learnable metrics produce noisy preferences that can compromise alignment - sometimes failing to optimize their own metric! 📉

Our solution: Use Flow-DPO (which only needs relative rankings!) with a surprisingly simple metric from 1982 - the Sampson epipolar error. Add rigorous data filtering (ensure meaningful gaps, remove static scenes) and you get a strong, principled approach to 3D consistency.

Why is enforcing 3D consistency so tricky? 1️⃣ Reconstruct → measure quality → use as reward: too slow 2️⃣ Manual labels → VLM → reward: noisy and expensive 3️⃣ Classical geometry metrics: often non-differentiable

Video diffusion models have made incredible progress and are increasingly used for 3D tasks like reconstruction and novel view synthesis. But there's still one major problem: they struggle with 3D consistency, producing geometric artifacts and unstable camera trajectories.

🎬 Epipolar Geometry Improves Video Generation Models ❗️ Excited to share our new work in collaboration with Fabian Manhardt, @fedassa and Christian Rupprecht! 🌐 Project Page: https://t.co/2rttYX9LPP Thread 🧵👇

📦 Released: 📄 Paper: https://t.co/NAzpEcXv26 🌐 Project: https://t.co/rD8HxMtGNb 🤗 Dataset: https://t.co/R6fhxd2gzZ 🎮 Demo: https://t.co/WoV7FkDa0t Diffusion models are knowledge engines. Time to tap into them. 🚀

Architecture: multi-mask decoder that predicts multiple valid hypotheses. Salient object detection is ambiguous: embrace it, don't average it away. Results: up to 50% error reduction on cross-dataset eval, SOTA on DIS & HR-SOD. Purely synthetic training matches real data. ✨

Generated 139K+ photorealistic samples with occlusions, complex backgrounds, diverse scenes. The pipeline scales with compute - generate as many as you need. The generation adapts to hard samples: evaluate performance → prioritize challenging categories → continuous improvement

We extract complementary information from three sources during generation: FLUX DiT features Concept attention maps DINO-v3 representations This extends the diffusion model to auto-generate masks alongside images, streamlining the entire data creation pipeline. 🎯

Modern diffusion models generate complex scenes with detailed lighting, occlusions, photorealistic textures. During generation, they encode massive amounts of spatial & semantic knowledge. Meanwhile, we're manually annotating data at 10 hours per sample. Something's broken. 🧵

❗️ Paper Release ❗️ S3OD: Towards Generalizable Salient Object Detection with Synthetic Data 🌐 Project Link:

In Donetsk region, russians killed more than 20 people — civilians who had gathered at that moment to receive their pensions. An air bomb was dropped on them. Earlier Trump said he had a good conversation with Vladimir Putin and that Putin definitely wants peace.

We built a system that detects highly complex objects NO vision model can find—introducing tool use for vision. 🧵 Say you wanted to detect for the @ycombinator logo in this image. (▶️ see step-by-step thinking) Using our solution, it's able to detect the logo perfectly,

Excited to share VMem: a novel memory mechanism for consistent video scene generation 🎞️✨ VMem evolves its understanding of scene geometry to retrieve the most relevant past frames, enabling long-term consistency 🌐 https://t.co/AHBj6j1ecE 🤗 https://t.co/FbUbJHWW4F 1/ 🧵

Many Congratulations to @jianyuan_wang, @MinghaoChen23, @n_karaev, Andrea Vedaldi, Christian Rupprecht and @davnov134 for winning the Best Paper Award @CVPR for "VGGT: Visual Geometry Grounded Transformer" 🥇🎉 🙌🙌 #CVPR2025!!!!!!

We are presenting Dual Point Maps as a #CVPR highlight tomorrow! Learn about our novel, data-efficient representation for 3D/4D deformable objects—an alternative to classical template shape models. 📍🕑 ExHall D, Poster #100, afternoon session 🌍 https://t.co/XclARf2SK7

🤖 Do you have a PhD, and want to push the frontier of computer vision and robotics? 🤖 The Visual Geometry Group (VGG) in Oxford is hiring a postdoc! PI: Dr. João Henriques. Deadline: 2 June at noon (UK). More details: https://t.co/XTAx9qgFEt

Sunday. Christian holiday before Easter. Barbaric brutal murder of civilians in Sumy. Russians did it intentionally. But never mind, @SteveWitkoff will thank Putin again and lick his bloody hands.