UDepFusion

Introduction

UDepFusion is an object-aware indoor RGB-D semantic mapping system based on YOLACT++ and Deep Learning based Depth Estimation model FRCN. The system is built upon the work of Elasticfusion¹ and Maskfusion².

UDepFusion is capable of

• Perform $geometric + semantic~segmentation$ with RGB-D (or raw RGB if D is missing) input image sequences.
• Assign model ids to the reconstructed models and perform tracking on individual models.
• Remove dynamic objects from the final semantic map by filtering

Additional Depth Estimation

Evaluation on TUM