PROSE: Training-Free Egocentric Scene Registration with Vision-Language Models

Name: PROSE: Training-Free Egocentric Scene Registration with Vision-Language Models
Uploaded: 2026-10-16T12:19:18.514318+00:00
Duration: 551 s
Description: PROSE is a training-free method for aligning two egocentric RGB scans taken at different times using a single pretrained Vision-Language Model (VLM). It builds 3D scene graphs for each scan and uses the VLM to match object instances across the scans. This approach outperforms geometric and learned scene-graph methods by focusing on object semantics, requiring no training or depth sensors.

PROSE is a training-free method for aligning two egocentric RGB scans taken at different times using a single pretrained Vision-Language Model (VLM). It builds 3D scene graphs for each scan and uses the VLM to match object instances across the scans. This approach outperforms geometric and learned scene-graph methods by focusing on object semantics, requiring no training or depth sensors.

Open the full episode