Uni-Fusion: Universal Continuous Mapping

InfoXVII - Robotics, Julius-Maximilians-Universität Würzburg

All you need, in one Uni. :D

1. Incremental Surface & Color Reconstruction

1.1 Mesh demonstration

We cut the ceiling for better visualization.

1.2 Surface reconstruction on ScanNet

Please zoom in the webpage and wipe the slidebar, BNV-Fusion's results contain lots of small particles in the wall.

BNV-Fusion
Uni-Fusion
BNV-Fusion
Uni-Fusion
BNV-Fusion
Uni-Fusion
BNV-Fusion
Uni-Fusion
BNV-Fusion
Uni-Fusion
BNV-Fusion
Uni-Fusion

1.3 Rendered Color on Replica

Fair continues mapping baseline NICE-SLAM does not reconstruct fine details. While Uni-Fusion is able to construct high quality result. We also compare with NeRF in paper to show the high postential but not demonstrate it here, as NeRF is task-specific to view-synthesis, which is not a fair baseline.

NICE-SLAM
Uni-Fusion
NICE-SLAM
Uni-Fusion
NICE-SLAM
Uni-Fusion
NICE-SLAM
Uni-Fusion

2. Style Transfer on 3D Canvas

3. Saliency Transfer on 3D Canvas

4. Scene Understanding

Uni-Fusion construct Surface Field of CLIP embedding.


4.1 Zero-shot Semantic Segmentation

4.2 Scene Understanding

5. All in one