3D Semantic MapNet: Building Maps for Multi-Object Re-Identification in 3D Environments
Konsep Inti
3D-SMNet is a model that re-identifies objects across different layouts in 3D environments, outperforming competitive baselines.
Abstrak
Abstract:
Study on 3D multi-object re-identification from embodied tours.
Creation of paired egocentric tours for object detection and re-identification.
Introduction:
Importance of re-identifying objects in changing environments for embodied agents.
Task Description:
Detecting and matching objects between layouts with changes in pose and context.
Approach:
Two-stage model: 3D object detector and matching module (SuperGlue).
Data Generation:
Use of Matterport3D scenes, YCB, and Google-scanned objects for dataset creation.
Experiments:
Training on generated episodes and zero-shot transfer to real-world rearrangement scenarios.
Related Work:
Comparison with existing methods like person re-ID and MOT.
Conclusion:
Significance of jointly training on real and simulated data for improved performance.
3D Semantic MapNet
Statistik
"On all datasets, we find 3D-SMNet outperforms competitive baselines."
"Jointly training on real and generated episodes can lead to significant improvements over training on real data alone."