LoGeR – 3D reconstruction from extremely long videos (DeepMind, UC Berkeley)
loger-project.github.io
March 10, 2026
4 min read
Summary
LoGeR enables dense 3D reconstruction of extremely long videos by processing them in chunks with a hybrid memory module, addressing quadratic complexity issues. It utilizes Sliding Window Attention for local alignment and Test-Time Training to maintain global consistency, effectively reducing drift over sequences of up to 19,000 frames.
Key Takeaways
- LoGeR utilizes a hybrid memory module to process long video streams in chunks, overcoming quadratic complexity bottlenecks in 3D reconstruction.
- The model achieves a 30.8% relative improvement over previous feedforward approaches on a 19,000-frame VBR dataset.
- LoGeR maintains strong geometric coherence and reduces scale drift over kilometer-scale trajectories without post-hoc optimization.
- The architecture combines Sliding Window Attention for local alignment and Test-Time Training for long-range consistency, enabling effective long-context reconstruction.
Community Sentiment
MixedPositives
- The potential for 3D reconstruction from long videos could revolutionize applications like street-view mapping, enhancing the accuracy of visual data collection.
- This research represents a significant advancement in video to 3D reconstruction, indicating rapid progress in the field of AI and computer vision.
Concerns
- There are serious concerns about the implications of this technology, particularly regarding its potential use in mass surveillance, which raises ethical questions.
- The accuracy of the model is questionable, as it may replicate lidar systems without providing reliable dimensional data, leading to potential inaccuracies.
Source
loger-project.github.io
Published
March 10, 2026
Reading Time
4 minutes
Relevance Score
54/100
🔥🔥🔥🔥🔥
Why It Matters
This page is optimized for focused reading: quick context up top, a clean summary block, and a direct path to the original source when you want the full story.