Multiple sites and room configurations.
Open research dataset · 2025Multispectral sensing · Indoor perception
IEEE Sensors Journal · First author
Indoor
MS.
IndoorMS: A Multispectral Dataset for Semantic Segmentation in Indoor Scene Understanding
17Buildings
19Semantic classes
82.38Best benchmark mF1
72.90Best benchmark mIoU
01 · Overview
Beyond
visible light.
A benchmark for asking what indoor perception can learn when sensing extends beyond conventional RGB imagery.
Indoor environments contain spectral information that ordinary colour cameras leave unobserved.
Indoor scene understanding is a critical computer-vision task, traditionally built around RGB semantic segmentation. IndoorMS introduces a dedicated multispectral benchmark to investigate the information available beyond the visible spectrum.
The dataset spans 17 buildings and diverse settings including meeting rooms, halls, lounges, offices, corridors and classrooms. Nineteen finely annotated semantic categories support robust evaluation of indoor segmentation models across different spatial layouts and spectral conditions.
Benchmark experiments with leading segmentation frameworks show that ConvNeXt-s with UperNet reaches an mF1 of 82.38 and an mIoU of 72.90, while also exposing class imbalance and domain gaps between RGB and multispectral representations as open research challenges.
02 · Dataset architecture
A new spectrum
for indoor scenes.
Fine pixel-level semantic annotations.
Rooms, halls, lounges, offices, corridors and classrooms.
Spectral observations beyond standard RGB sensing.
Designed specifically for multispectral semantic scene understanding rather than repurposed from RGB collections.
A testbed for representation learning, multimodal transfer and long-tailed semantic prediction.
03 · Representation
One room.
Three ways to see it.
04 · Annotation gallery
Indoor detail,
densely labelled.
06 · Access
Use the
dataset.
Dataset, code and documentation.
Visit the official repository for data access, benchmark resources and project updates.
IEEE Sensors Journal.
Read the peer-reviewed paper for acquisition details, benchmark protocols, results and open research questions.
07 · Citation
Build on
this work.
@article{10965893,
author = {Zhu, Qinfeng and Xiao, Jingjing and Fan, Lei},
journal = {IEEE Sensors Journal},
title = {IndoorMS: A Multispectral Dataset for Semantic Segmentation in Indoor Scene Understanding},
year = {2025},
doi = {10.1109/JSEN.2025.3559348}
}