Emily Steiner

PhD Candidate @ Stanford | Gradient 𝚫 Spaces Group | Researching 3D/4D Scene Understanding

[Alyssa Unell @AlyssaUnell]

1/ Everyone's deploying AI nowadays — but how do we evaluate these systems once they're actually in use?
Excited to share our team's new preprint: "Deployment-Centered Evaluation: Predicting Query-Level Rejection Risk in a Clinical LLM System." 🧵

[Iro Armeni @ir0armeni]

Three reasons to find Gradient Spaces at @CVPR this week 👇

🧩 Scenes don't sit still; furniture moves, rooms change. 𝗥𝗲𝗦𝗰𝗲𝗻𝗲𝟰𝗗 keeps semantic instance segmentation temporally consistent across evolving indoor 3D scenes.
📍 Fri 4–6 PM · ExHall A & F · #329
🌐 https://t.co/EDKpdAdS2F
🤹 w/ @easteine, @jianhao75895505, Henry Howard-Jenkins, Christopher Xie
🔖 https://t.co/1ijL8bNQtD

🌆 What if your video generator understood geometry? 𝗚𝗮𝘂𝘀𝘀𝗙𝘂𝘀𝗶𝗼𝗻 does exactly that — for sharper 3D reconstruction and 3DGS in the wild.
📍 Sat 11:45 AM–1:45 PM · ExHall F · #111
🌐 https://t.co/X1eju438BW
🤹 w/ @liyuan_zz, Manjunath Narayana, Michal Stary, Will Hutchcroft, @GordonWetzstein
🔖 https://t.co/hrRDBZgeBK

🎯 Pose estimation that doesn't flinch at messy real-world motion. 𝗪𝗶𝗹𝗱𝗣𝗼𝘀𝗲 is a single approach for trajectories of any length and dynamics.
📍Sun 11:45 AM–1:45 PM · ExHall F · #26
🌐https://t.co/RvYvL0OeIs
🤹w/ @jianhao75895505, @liyuan_zz, @zhuzihan2000
🔖 https://t.co/kFAK4AGnNH

Come find us at the posters!

#CVPR2026 #3DVision #4DSceneUnderstanding #DynamicScenes #3DReconstruction #PoseEstimation #GradientSpaces

[Iro Armeni @ir0armeni]

3DGS produces artifacts; they can be fixed, but why do it in 2D?
GaussFusion, @CVPR, conditions a video model on a multi-modal Gaussian Primitive buffer so it reasons about 3D structure, not just pixels
Handles diverse 3DGS pipelines & their artifacts, efficient via distillation

[Emily Steiner @easteine]

💼 Work done @Stanford w/ @jianhao75895505 , Meta @RealityLabs (Henry Howard-Jenkins, Chris Xie), and @ir0armeni ✨

@CVPR

📰 https://t.co/4YPyOta24N
🌐 https://t.co/OGpmToDSri

#ComputerVision #3DPerception #CVPR2026 #GradientSpaces

[Emily Steiner @easteine]

🚀 New paper at #CVPR2026: https://t.co/63flacDArE

Indoor scenes naturally evolve over long horizons—furniture and objects move, deform, appear, and disappear. How do we maintain consistent instance identities across days, months, or years? 🏠📅

Introducing ReScene4D: temporally consistent 4D semantic instance segmentation for sparsely observed indoor scenes 🔀

[Emily Steiner @easteine]

💡 Key insight: Re-observing scenes improves standard per-scene 3D segmentation—even if there are changes!

Multiple observations act as augmentation, consolidating evidence from partial scans, incomplete views, and occlusions. Result: we outperform 3D-only baselines on per-scan metrics by fusing temporal information 📊

[Iro Armeni @ir0armeni]

The Nothing Stands Still Challenge is back and due on May 7th!
💰 Kept the cash prize but created new data that are extra challenging.
🏗️ Goal: create spatiotemporal maps of drastically changing environments. What a better fit than construction sites?
🌐https://t.co/PlYitDVKSt