CoPE-VideoLM: Leveraging Codec Primitives for Efficient Video Language Modeling
Uses video codec primitives instead of dense per-frame embeddings, cutting time-to-first-token by up to 86% and tokens by up to 93%.
About Me
I am a PhD candidate (2024–) in the Gradient Spaces Group at Stanford University, advised by Prof. Iro Armeni, and part of the Stanford Vision Lab. During my PhD, I have interned at Waymo Perception (Summer ’26), working on streaming temporal reasoning, and at Microsoft Spatial AI Lab (Summer ’25), working on efficient video tokenization. I am partly supported by a Google XR grant on 4D understanding.
My research focuses on spatial intelligence and multimodal video understanding. I build efficient visual representations that capture space and time, align real-world 3D scenes across modalities, and enable controllable built-environment generation.
Before PhD, I received my M.Sc. in Computer Science from ETH Zurich, advised by Prof. Marc Pollefeys, during which I interned at Qualcomm XR working on real-time SLAM. Earlier, I was a computer vision research engineer at Mercedes-Benz R&D and spent a year in Prof. Vincent Lepetit's lab at TU Graz working on hand-object pose estimation.
Interests
- Spatial Intelligence
- Multimodal Video Understanding
- Visual Representation Learning
News
I joined Waymo Perception (Mountain View) as a research intern, working on streaming temporal reasoning.
Recognized as an Outstanding Reviewer at CVPR 2026. Very proud of this!
Invited talks on “Codec Primitives for Efficient Video Understanding” at Google DeepMind and Valeo AI in Paris.
CoPE-VideoLM is out: efficient codec-aware tokenization for video understanding.
Invited talk on GuideFlow3D at Voxel51 Best of NeurIPS.
GuideFlow3D was accepted to NeurIPS 2025. See you in San Diego!
Invited talks on “Scalable Cross-Modal 3D Scene Understanding” at Google XR Research and Imagine Lab.
I joined Microsoft Spatial AI Lab (Zurich) as a research intern, working on efficient video tokenization.
Show moreShow less
CrossOver was accepted to CVPR 2025 as a Highlight. See you in Nashville!
Career update I joined Stanford University as a PhD student in Computer Vision.
I joined Qualcomm XR (Amsterdam) as a research intern, working on real-time SLAM for extended reality.
SGAligner was accepted to ICCV 2023. See you in Paris!
I started my MSc in Computer Science at ETH Zurich.
Keypoint Transformer was accepted to CVPR 2022 as an Oral. See you in New Orleans!
I joined Mercedes-Benz R&D as a Computer Vision Research Engineer.
Monte Carlo Scene Search was accepted to CVPR 2021.
General 3D Room Layout from a Single View by Render-and-Compare was accepted to ECCV 2020.
I started as a Research Assistant at IVC, TU Graz with Prof. Vincent Lepetit, supported by a Qualcomm fellowship.
Invited Talks
-
Codec Primitives for Efficient Video Understanding SlidesApr 2026Google DeepMind, London / ParisHost(s): Guillaume Le Moing and Tengda HanMar 2026Valeo AI, ParisHost: Andrei Bursuc
-
Optimization-Guided Rectified Flow for Appearance Transfer SlidesVoxel51 TalkJan 2026Voxel51 Best of NeurIPSHost: Jimmy GuerreroNov 2025SVL Vision Seminar (Student Talks), StanfordHost: Prof. Jiajun Wu
-
Scalable Cross-Modal 3D Scene Understanding SlidesNov 2025Google XR Research, MunichHost: Johanna WaldJul 2025Imagine Lab, École des Ponts, ParisHost(s): Raphael Baena and Prof. Vincent Lepetit
Research
Visual Representation Learning
Efficient visual encoders for spatial and video understanding.
Controllable Generation
Guiding pre-trained generative models to control shape and appearance.
Publications
Google Scholar* Equal contribution · † Equal supervision
2026
SpaceFlow: Locally Controllable 3D Generation
Training-free 3D generation from geometric primitives, with per-part control over shape adherence and part-specific text or image appearance cues.
FAMOS: Feed-Forward 3D Articulation Modeling from Sparse Observations
A feed-forward model that predicts movable parts and joint parameters of articulated objects from a sparse, unordered set of partial point clouds.
2025
SGAligner++: Cross-Modal Language-Aided 3D Scene Graph Alignment
Extends SGAligner with open-vocabulary cues and learned joint embeddings to align 3D scene graphs across modalities, even under low overlap and sensor noise.
CrossOver: 3D Scene Cross-Modal Alignment
Cross-modal alignment method for 3D scenes that learns a modality-agnostic embedding space, enabling scene-level alignment without semantic annotations.
Featured inOpen Robotics
2023
SGAligner: 3D Scene Alignment with Scene Graphs
3D Scene Graph Alignment robust to in-the-wild scenarios powering point cloud registration and map integration.
2022
Keypoint Transformer: Solving Joint Identification in Challenging Hands and Object Interactions for Accurate 3D Pose Estimation
Efficient network for two-hand and object pose estimation in complex interactions, paired with the H2O-3D dataset of two-hand interaction with YCB objects.
2021
Monte Carlo Scene Search for 3D Scene Understanding
Monte-Carlo Tree Search (MCTS) based analysis-by-synthesis method to recover complete scene (3D layout+objects) from a noisy RGB-D scan.
2020
General 3D Room Layout from a Single View by Render-and-Compare
3D layout estimation from a single perspective view, to recover complex non-cuboid layouts by solving a constrained discrete optimization problem.
Background
Experience
- Waymo PerceptionPhD Research Intern · Mountain ViewSummer 2026
- Microsoft Spatial AI LabResearch Scientist Intern · ZurichSummer 2025
- Qualcomm XRResearch Engineer Intern · Amsterdam2023
- Mercedes-Benz R&DComputer Vision Research Engineer · Bangalore2021–2022
- IVC, TU GrazResearch Engineer · Graz2020–2021
Education
- Stanford UniversityPh.D. in Computer Vision2024–
- ETH ZurichM.Sc. in Computer Science2022–2024
- Manipal Institute of TechnologyB.Tech. in Information Technology2016–2020
Supervision
- M.Sc. Robotics student at ETH Zurich, now CS PhD in Prof. Siyu Tang’s group at ETH Zurich
- Summer Research Fellow at INSAIT, now MPhil student at the University of Cambridge
- M.Sc. Computational Mechanics student at TUM, now CS PhD with Prof. Theodora Kontogianni
Academic Services
- Workshop OrganizationCV4AEC @ CVPR 2023, 2024, 2026
- Conference ReviewerCVPR, ICCV, ECCV, NeurIPS, ICRA · Outstanding Reviewer, CVPR 2026
- TeachingHead TA, Computer Vision for the Built Environment, Stanford · Winter 2025, 2026
Get in touch
Let’s talk research.
I’m always open to research collaborations. If you’re around the Bay Area, let’s grab a coffee!