IGGT4D proposes a streaming Transformer for real-time 4D scene understanding
liuziwei7 · x · 2026-07-27
IGGT4D introduces a streaming Transformer for real-time 4D scene understanding
The paper proposes IGGT4D, a Streaming 4D Instance-Grounded Geometry Transformer for online scene understanding in dynamic videos.
- It incrementally processes video frames to build a unified 4D representation that captures camera motion, geometry, and object identity.
- The goal is to handle real-world scenes where objects move, disappear, and reappear, while keeping instance-level consistency over time.
- The paper also introduces InsScene4D-147K, a large-scale dataset with geometry-guided instance masks for training and evaluation.
- The authors say the representation supports downstream tasks including instance spatial tracking, open-vocabulary segmentation, and scene grounding.
More from Embodied
- Prototype muscle topology works, and the next version is already getting smaller — _Stocko_ · 2026-07-27
- Three AI devices move from the pocket to the body: camera, mic, and tongue controller — APPSO · 2026-07-27
- A New York school district will put a $57,000 humanoid robot into classrooms this fall — CyberRobooo · 2026-07-27
- YC says physical AI now feels like GPT-2 and DALL·E did a few years ago — ycombinator · 2026-07-27
- Shenzhen adds 1,227 ultra-fast chargers as Chengdu robot makers ship one unit every 12 minutes — 创业邦 · 2026-07-27
- Maker Revisits Gear Train for a 52-Weeks Robot Build — _Stocko_ · 2026-07-27