←
Marwan Taher79 posts
Marwan Taher avatar

Marwan Taher

@marwan_ptr

PhD student at the Dyson Robotics Lab @ Imperial College London.

◷ Joined April 2011▧ 10 Photos and videos
Ivan Kapelyukh 🇺🇦 avatar
⟳ Marwan Taher retweeted
Ivan Kapelyukh 🇺🇦@IvanKapelyukh·Sep 29

Simify: use generative real2sim to build a simulator, then perform spatial reasoning using evolutionary search. Presented by @YafeiHuCMU at IROS! arxiv.org/abs/2609.33982

Kirill Mazur avatar
⟳ Marwan Taher retweeted
Kirill Mazur@makezur·Sep 14

AgentSTAR is an agentic method for both reconstruction and tracking of articulated objects from monocular videos. We can now track through rapid motion, severe occlusion and thin objects! 

There’s no spoon but we can still track it 🥄

0:54
Marwan Taher avatar
Marwan Taher@marwan_ptr·Jun 6

Wanna see a real-time demo of KV-Tracker? Stop by in ExHall F today 11:45 - 13:45! #CVPR2026

0:24
Kirill Mazur avatar
⟳ Marwan Taher retweeted
Kirill Mazur@makezur·Jun 5

How to build complete 4D reconstructions from videos? Swing by 4DPM oral presentation today: Mile High Ballroom 3A - 4A, 13:00 #CVPR2026

0:07
Kirill Mazur avatar
⟳ Marwan Taher retweeted
Kirill Mazur@makezur·Apr 16

4DPM got selected for an oral presentation at #CVPR2026 🤠

Introducing 4D Primitive-Mâché (4DPM), a new method for replayable 4D reconstruction from monocular videos. We split dynamic scenes into 3D primitives and recover their motion. 4DPM can infer object positions even after they leave view. Joint work with @marwan_ptr @AjdDavison

0:30
Marwan Taher avatar
Marwan Taher@marwan_ptr·19 Dec 2025

KV-Tracker enables object-level reconstruction and tracking when provided with an object mask. The KV-cache can be saved and used later without any special initialisation procedure.

0:34
Marwan Taher avatar
Marwan Taher@marwan_ptr·19 Dec 2025

KV-Tracker is suitable for online reconstruction and real-time tracking, capable of maintaining 30 FPS for up to ~45 keyframes. Our caching approach provides a ~10x boost in inference speed!

0:40
Marwan Taher avatar
Marwan Taher@marwan_ptr·19 Dec 2025

Our online pipeline is simple, we run π³ on a set of keyframes, recovering the underlying geometry and generating a KV-cache that can be used for tracking. Effectively the KV-cache is our underlying implicit scene representation.

Post attachment
Marwan Taher avatar
Marwan Taher@marwan_ptr·19 Dec 2025

Models like VGGT, π³ and Depth Anything 3, use global self-attention blocks to exchange information between the input views. To overcome this quadratic complexity, we cache key and value pairs for a set of keyframes and use them for inference with latest frame.

0:54
Marwan Taher avatar
Marwan Taher@marwan_ptr·18 Dec 2025

Per-frame geometry from π³ is split into primitives via segmentation and tracked over time using dense 2D point tracks. With a compact per primitive pose, geometry is densely aligned, stitching primitives to create a complete reconstruction of the observed scene components.

Introducing 4D Primitive-Mâché (4DPM), a new method for replayable 4D reconstruction from monocular videos. We split dynamic scenes into 3D primitives and recover their motion. 4DPM can infer object positions even after they leave view. Joint work with @marwan_ptr @AjdDavison

0:30