R4DSG: Relative 4D Scene Graph Memory for Object-Centric Question Answering in Long Egocentric Video
The paper proposes a compact object memory for long wearable-camera video that tracks items by relative change, not a reconstructed world map.
R4DSG separates stable anchors from dynamic objects, keeps object identity across frames, and stores state changes as anchor-relative transitions. The authors say this is meant to answer object-centric questions such as where something moved, when it changed state, or why it was relocated. On a 255-question EgoLifeQA subset, it reports a 6.7-point overall gain over EgoRAG-Text, including 12.5 points on “when” questions. ArXiv · AI/CL/LG's note
R4DSG separates stable anchors from dynamic objects, keeps object identity across frames, and stores state changes as anchor-relative transitions. The authors say this is meant to answer object-centric questions such as where something moved, when it changed state, or why it was relocated. On a 255-question EgoLifeQA subset, it reports a 6.7-point overall gain over EgoRAG-Text, including 12.5 points on “when” questions. ArXiv · AI/CL/LG's note
score 5