3D Scene Segmentation: A Comprehensive Survey and Open Problems

This paper presents a detailed review of recent advancements in 3D indoor scene segmentation driven by deep learning techniques. It provides an overview of existing segmentation models, examines various data representations, data collection methods, augmentation techniques, and available datasets. A comparative analysis of loss functions and overview of evaluation metrics is conducted to highlight their impact on segmentation performance. Unlike previous surveys, this work introduces a new classification of data augmentation techniques and proposes two novel classification approaches for 3D instance and semantic segmentation. Furthermore, it unifies 3D semantic instance segmentation and 3D panoptic segmentation within an existing framework. The paper also identifies key challenges and open research directions, providing insights into future advancements in the field.