Cross-Dataset Semantic Segmentation Performance Analysis: Unifying NIST Point Cloud City Datasets for 3D Deep Learning

Published

February 23, 2026

Author(s)

Alexander Dimopoulos, Joseph Grasso

Abstract

This study analyzes semantic segmentation performance across heterogeneously labeled point cloud datasets focused on emergency response applications. Using NIST's Point Cloud City dataset from Enfield and Memphis collections, we investigate the fundamental challenges of unifying differently labeled 3D data for public safety. Our methodology develops a graded schema compatible with state-of-the-art KPConv (Kernel Point Convolution) [] architecture and evaluates performance using IoU metrics across safety-critical features. Results reveal significant performance variability, with geometrically large and similar objects (like stairs and windows) performing exceptionally well, while smaller safety-critical features crucial for first responders [] exhibited poor recognition rates. Severe class imbalances further undermined model reliability for minority classes essential to emergency response. These findings highlight fundamental limitations in current point cloud approaches for detecting smaller objects of high public safety value, while demonstrating potential for creating navigational understanding through larger geometric features. Our work identifies three critical challenges []: insufficient labeled data despite large point clouds, inefficiencies in unifying class labels across heterogeneous datasets resulting in lost information, and the need to align with emerging industry standardization efforts. Potential solutions include automating the labeling process through 2D visual language models with segmentation capabilities, or implementing prompt-based approaches for multi-dataset learning to overcome negative transfer effects. We conclude that reliable point cloud semantic segmentation for public safety applications requires both standardized annotation protocols and automated labeling techniques to address the complexity and scale of 3D environments, particularly for pre-incident planning systems generated from lidar scans.

Conference Dates

September 23-25, 2025

Conference Location

Orlando, FL, US

Conference Title

2025 IEEE World Forum on Public Safety Technology (WF-PST)

Pub Type

Conferences

Citation

Dimopoulos, A. and Grasso, J. (2026), Cross-Dataset Semantic Segmentation Performance Analysis: Unifying NIST Point Cloud City Datasets for 3D Deep Learning, 2025 IEEE World Forum on Public Safety Technology (WF-PST), Orlando, FL, US (Accessed February 24, 2026)