Few-Shot Segmentation of 3D Point Clouds Under Real-World Distributional Shifts in Railroad Infrastructure
Industrial railway monitoring systems require precise understanding of 3D scenes, typically achieved using deep learning models for 3D point cloud segmentation. However, real-world applications demand these models to rapidly adapt to infrastructure upgrades and diverse environmental conditions acros...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-02-01
|
| Series: | Sensors |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1424-8220/25/4/1072 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Industrial railway monitoring systems require precise understanding of 3D scenes, typically achieved using deep learning models for 3D point cloud segmentation. However, real-world applications demand these models to rapidly adapt to infrastructure upgrades and diverse environmental conditions across regions. Conventional deep learning models, which rely on large-scale annotated datasets for training and are evaluated on test sets that are drawn independently and identically from the training distribution, often fail to account for such real-world changes, leading to overestimated model performance. Recent advancements in few-shot learning, which aim to develop generalizable models with minimal annotations, have shown promise. Motivated by this potential, the paper investigates the application of few-shot learning to railway monitoring by formalizing three types of distributional shifts that are commonly encountered in such systems: (a) in-domain shifts caused by sensor noise, (b) in-domain out-of-distribution shifts arising from infrastructure changes, and (c) cross-domain out-of-distribution shifts driven by geographical variations. A systematic evaluation of few-shot learning’s adaptability to these shifts is conducted using three performance metrics and a predictive uncertainty estimation metric. Extensive experimentation demonstrates that few-shot learning outperforms fine-tuning and maintains strong generalization under in-domain shifts with only ~1% performance deviation. However, it experiences a significant drop in performance under both in-domain and cross-domain out-of-distribution shifts, pronounced when dealing with previously unseen infrastructure classes. Additionally, we show that incorporating predictive uncertainty estimation enhances few-shot learning applicability by quantifying the model’s sensitivity to distributional shifts, offering valuable insights into the model’s reliability for safety-critical applications. |
|---|---|
| ISSN: | 1424-8220 |