IDA-Net: Intensity-distribution aware networks for semantic segmentation of 3D MLS point clouds in indoor corridor environments

Zhipeng Luo, Pengxin Chen, Wenzhong Shi, Jonathan Li

Abstract

Semantic segmentation of 3D point clouds from Mobile Laser Scanning (MLS) often struggles with the "modality gap" between geometric coordinates and laser intensity. Traditional methods treat these as a single concatenated vector, often leading to the suppression of intensity-based material information. This paper introduces IDA-Net, a framework featuring an Intensity-Distribution Aware (IDA) descriptor and a two-stage embedding strategy. By analyzing intensity histograms and using them to guide training and refine testing, IDA-Net significantly improves segmentation accuracy in indoor corridors, achieving over 90% mIoU.

Introduction

In 3D Mobile Laser Scanning (MLS), each point typically possesses four primary attributes: spatial coordinates ($x, y, z$) and laser reflection intensity ($I$). While most deep learning models treat these as a unified 4D vector, this approach overlooks the fundamental "modality gap." Coordinates define geometry (spatial location), whereas intensity reflects material properties (surface reflectance).

In complex indoor environments like corridors, objects such as wooden doors and concrete walls may share similar geometric profiles but exhibit distinct reflective signatures. Blindly stacking these modalities often causes neural networks to prioritize modality-agnostic features, effectively drowning out the unique signals provided by intensity.

Methodology: The IDA-Net Framework

To address the modality gap, IDA-Net introduces two core components that process geometry and intensity with specialized logic. The architecture separates the point cloud into geometric features and intensity-based masks, which are then integrated in a two-stage process.

The two-stage IDA-Net architecture showing the dual-path processing of geometry and intensity.

1. The IDA Descriptor

The Intensity-Distribution Aware (IDA) descriptor avoids feeding raw intensity directly into the backbone. Instead, it utilizes "Multi-peak searching" (MPS) to analyze scene histograms across the x, y, and z axes. This module identifies clusters of points with similar intensity patterns to generate a multi-channel Intensity-Distribution (ID) mask. This mask represents the probability of a point belonging to a specific category based solely on its reflective "fingerprint."

Detailed workflow of the Intensity-Distribution Aware (IDA) descriptor.

2. Two-Stage Embedding Strategy

The ID mask is integrated into the network through a two-stage process:

  • Training Stage (Guiding): The ID mask is used to calculate a specialized "ID loss," forcing the network to learn features consistent with real-world intensity distributions.
  • Testing Stage (Refining): During inference, the ID mask serves as a "confidence filter," multiplying initial predictions to suppress noise and correct geometry-based errors using intensity logic.

Experimental Setup and Results

The model was evaluated on two real-world indoor corridor datasets:

  • Area 1: A 250-meter corridor with five categories: Floor, Ceiling, Wall, Door, and Window frame.
  • Area 2: A 300-meter corridor used to test adaptability across different architectural layouts.

The quantitative improvements were significant. Using KPConv as a backbone, the mean Intersection over Union (mIoU) rose from 80.76% to 90.58% in Area 1, and from 81.58% to 88.94% in Area 2. Qualitatively, IDA-Net excelled at distinguishing "inlaid" objects, such as doors flush with walls, which are frequently misclassified by geometry-only models.

Key Insights

  • Intensity as a Prior: Treating intensity as a source of prior knowledge (via the ID loss) allows material properties to influence model weights without interfering with geometric feature extraction.
  • Backbone Agnostic: IDA-Net functions as a modular framework. Tests across four different backbones (PointNet, PointNet++, PointConv, and KPConv) showed consistent performance gains regardless of the underlying architecture.
  • Structured Information: The success of the IDA descriptor demonstrates that traditional statistical tools like histograms can provide highly interpretable and structured information that aids neural network convergence.

Conclusion and Future Work

While highly effective in structured corridor environments, the current IDA descriptor relies on manual design for histogram binning. Future research aims to develop an automatic, data-driven IDA descriptor. By employing a sub-network to learn optimal intensity patterns, the system could be made robust enough for autonomous driving in complex, unpredictable urban outdoor scenes.

Summary updated

BibTeX
@article{luo2022ida,
  author = {Zhipeng Luo and Pengxin Chen and Wenzhong Shi and Jonathan Li},
  title = {{IDA-Net: Intensity-distribution aware networks for semantic segmentation of 3D MLS point clouds in indoor corridor environments}},
  journal = {Int. J. Appl. Earth Obs. Geoinformation},
  year = {2022},
  doi = {10.1016/j.jag.2022.102904},
}
Figure