Ledger: Persistent 3D Object Memory for Egocentric Videos
What it is
This paper introduces Ledger, a system that builds a persistent 3D object memory from egocentric videos by tracking object locations, histories, and contextual descriptions. It clusters observations by resting locations to filter noise and retains objects after they leave the view, enabling spatial queries without accessing raw video data.
The problem it solves
Existing systems struggle to maintain persistent, accurate 3D object memories from egocentric videos, often failing to track objects after they leave the field of view or suffering from localization noise.
What you could use it for
- Answering spatial questions about object locations in past egocentric recordings
- Tracking object movement and history in embodied AI assistants
- Improving accuracy on egocentric video benchmarks like HD-EPIC and UCS-Bench
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.