NeRF-SOS: Any-View Self-supervised Object Segmentation on Complex Scenes


1 The University of Texas at Austin
2 The University of Texas at Austin
3 The University of Texas at Austin

The University of Texas at Austin

The University of Texas at Austin

The University of Texas at Austin

The University of Texas at Austin

The University of Texas at Austin

The University of Texas at Austin

Paper
</Code>

tl, dr

We propose a novel collaborative contrastive loss for NeRF to segment objects in complex real-world scenes, without any annotation.

We augment NeRF models by appending a parallel segmentation branch to predict point-wise implicit segmentation feature. We propose to update the segmentation feature field using a collaborative loss in both appearance and geometry levels. During inference, a clustering operation (e.g., K-means) is used to generate object masks, based on the rendered feature field.

Citation

@misc{https://doi.org/10.48550/arxiv.2209.08776,
doi = {10.48550/ARXIV.2209.08776},
url = {https://arxiv.org/abs/2209.08776},
author = {Fan, Zhiwen and Wang, Peihao and Jiang, Yifan and Gong, Xinyu and Xu, Dejia and Wang, Zhangyang},
keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
title = {NeRF-SOS: Any-View Self-supervised Object Segmentation on Complex Scenes},
publisher = {arXiv},
year = {2022}}