MJU-Glass: RGB-D Benchmark for Glass Segmentation

• MJU-Glass is an RGB-D dataset designed for precise binary segmentation of glass surfaces under challenging conditions.
• It uses dual-view data from a mobile robot to capture indoor and outdoor scenes, addressing issues like reflections, occlusions, and weak boundaries.
• Adaptive fusion techniques, such as the Weighted Feature Fusion module, significantly improve glass boundary detection and overall segmentation performance.

MJU-Glass is an RGB-D dataset for glass surface segmentation introduced for service-robot perception. It contains 4,900 RGB-D image pairs collected by a wheeled mobile robotic platform equipped with two Orbbec Gemini Pro cameras, and it is intended to benchmark binary pixel-wise segmentation of glass against background in indoor and outdoor scenes, including extreme environments. The dataset is motivated by the fact that glass surfaces are difficult to segment because of transparency, specular reflections / highlights, lack of texture, background ambiguity, light refraction and reflection, occlusion by or visibility of objects behind the glass, and unclear or weak boundaries; in depth images, glass may additionally produce missing depth values, distorted depth readings, and irregular depth discontinuities (Lin et al., 3 Aug 2025).

1. Task definition and design objective

MJU-Glass was created for glass surface segmentation in service-robot settings rather than for generic transparent-object recognition. Its target is the robotic problem of identifying glass surfaces such as glass doors, windows, tabletops, and partitions from the viewpoint of a mobile robot operating in real environments. The intended applications stated for the dataset include safe navigation, obstacle avoidance, and improved SLAM/environmental modeling around transparent obstacles (Lin et al., 3 Aug 2025).

The dataset’s design reflects a specific failure mode of conventional vision systems. In RGB imagery, glass often resembles the background or the objects visible behind it, whereas depth sensing around glass can be unstable or incomplete. This makes MJU-Glass explicitly an RGB-D benchmark rather than an RGB-only resource. A plausible implication is that the dataset was designed to test whether modality fusion can recover glass boundaries and occupancy cues that neither modality provides robustly in isolation.

2. Robotic acquisition platform and sensor geometry

The data were collected using a wheeled mobile robotic platform with dimensions $$565.2 \text{ mm} \times 774.5 \text{ mm} \times 180 \text{ mm}$$, maximum speed $0.46 \text{ m/s}$, and onboard computer NVIDIA Jetson TX1. The paper states that the setup was derived from an actual robotic platform used in commercial products developed by Fujian Hantewin Intelligent Technology Co., Ltd. (Lin et al., 3 Aug 2025).

Two Orbbec Gemini Pro RGB-D cameras were mounted on the robot in a two-view configuration intended to mimic realistic service-robot viewpoints:

• Camera 1: height 200 mm, orientation $0^\circ$ relative to the horizontal ground, intended for close-range or ground-level data.
• Camera 2: height 950 mm, orientation $-45^\circ$ relative to the horizon, intended to observe the ground and nearby objects with larger projected size, which the paper describes as helpful for object detection and spatial mapping.

The sensor specifications reported in the paper are detailed but internally inconsistent for depth resolution. For the depth sensing module, the paper reports working range 0.25 m to 2.5 m, precision $\pm 5$ mm at 1 m, frame rate 60 fps, and field of view $67.9^\circ(H) \times 45.3^\circ(V)$. One passage states that original depth images were $1280 \times 800$, while another states that the depth module captures $640 \times 400$ at 60 fps. For the RGB module, the paper reports $640 \times 480$ resolution, 60 fps, and field of view $71.5^\circ(H) \times 56.7^\circ(V)$. The reliable benchmark-side statement is that, after preprocessing, depth images were aligned to RGB and resized to $0.46 \text{ m/s}$0 (Lin et al., 3 Aug 2025).

The paper further states that RGB-D pairs were collected from a physical robot navigating diverse environments, which strongly suggests acquisition during robot movement. However, whether every frame was captured while moving is not specified.

3. Dataset composition and annotation semantics

The dataset consists of 4,900 images, described as RGB-D image pairs. Its semantic task is binary segmentation with the mask formalized as

$0.46 \text{ m/s}$1

where $0.46 \text{ m/s}$2 denotes background and $0.46 \text{ m/s}$3 denotes glass. The annotation granularity is therefore pixel-wise semantic segmentation of glass surfaces against background (Lin et al., 3 Aug 2025).

The paper makes clear what the dataset supports and what it does not explicitly define. It supports binary pixel-wise segmentation and reports boundary-sensitive evaluation derived from these masks, but it does not mention instance labels, panoptic labels, object IDs, polygon annotations, or separately stored boundary labels. Likewise, the annotation procedure itself is not described: the paper does not specify whether masks were drawn manually, what tool was used, whether multiple annotators were involved, whether quality control was applied, or how ambiguous transparent regions were handled (Lin et al., 3 Aug 2025).

Scene coverage is described qualitatively rather than by a detailed taxonomy. MJU-Glass includes indoor and outdoor scenes, including extreme environments, and emphasizes difficult conditions such as bright lighting, poor lighting, high transparency, complex environments, reflections, occlusions, and interference from objects behind the glass. The paper identifies four difficult-case categories for a dedicated robustness subset: bright lighting, poor lighting, high transparency, and complex environments (Lin et al., 3 Aug 2025).

A precise training/validation/test split for the full set of 4,900 images is not specified. What is specified is a difficult test set of 400 images, constructed by sorting test results of common models and selecting the 400 images with the lowest mIoU from DeepLabv3+ and PSPNet. The paper therefore distinguishes an All benchmark and a Difficult benchmark, but it does not provide full partition counts for training, validation, and ordinary testing (Lin et al., 3 Aug 2025).

4. Benchmark protocol and reported performance

MJU-Glass is used as a benchmark for binary semantic segmentation with three principal evaluation metrics: IoU for glass, mIoU, and boundary IoU (bIoU). The paper states that boundary IoU is evaluated within a 5-pixel range, reflecting the importance of accurately localizing difficult glass boundaries rather than merely segmenting large easy background regions (Lin et al., 3 Aug 2025).

The main baselines are DeepLabv3+ and PSPNet, both with ResNet-50 backbone and ImageNet-pretrained weights, evaluated in three settings: RGB-only, simple RGB-D feature fusion by concatenation, and RGB-D with the proposed Weighted Feature Fusion (WFF). The paper does not specify optimizer, learning rate, number of epochs, batch size, augmentation, validation protocol, or test-time augmentation for these benchmarks (Lin et al., 3 Aug 2025).

The reported results are as follows:

Model / setting	All: IoU / mIoU / bIoU	Difficult: IoU / mIoU / bIoU
DeepLabv3+ RGB only	89.61 / 93.17 / 38.09	88.74 / 92.12 / 38.10
DeepLabv3+ concat	89.76 / 93.36 / 39.91	88.74 / 92.21 / 38.56
DeepLabv3+ WFF	90.12 / 93.61 / 40.35	88.80 / 92.36 / 39.44
PSPNet RGB only	86.74 / 91.43 / 25.33	85.90 / 90.31 / 25.60
PSPNet concat	90.21 / 93.67 / 32.66	88.74 / 91.49 / 30.17
PSPNet WFF	90.47 / 93.84 / 33.69	89.39 / 92.43 / 33.09

These numbers show a consistent trend: RGB-D fusion improves over RGB-only baselines, and WFF yields the best overall performance on both All and Difficult subsets. The gains are especially pronounced in bIoU, which the authors interpret as evidence that adaptive fusion helps most at glass contours and other boundary-dominated regions. The paper highlights a 7.49% improvement in bIoU for PSPNet when integrating WFF; the table values for the All setting are 25.33 for RGB-only and 33.69 for WFF, which is an 8.36-point absolute difference (Lin et al., 3 Aug 2025).

MJU-Glass was introduced together with the Weighted Feature Fusion module, whose central fusion equation is

$0.46 \text{ m/s}$4

where $0.46 \text{ m/s}$5 and $0.46 \text{ m/s}$6 are RGB and depth feature maps and the weights are normalized channel-wise. This is method context rather than a dataset property, but it is integral to how the benchmark was framed: MJU-Glass is used to measure whether adaptive RGB-D weighting improves segmentation robustness under service-robot conditions (Lin et al., 3 Aug 2025).

5. Position within the glass-perception literature

The paper explicitly compares MJU-Glass with prior RGB-D glass-related datasets and presents it as more tightly aligned with real-world service-robot glass segmentation than narrower or more curated alternatives (Lin et al., 3 Aug 2025).

Dataset	Images	Scene/classes summary
MJU-Glass	4,900	indoor and outdoor scenes, including extreme environments; classes: glass, background
RGBD-Mirror	3,049	various indoor environments with glasses; classes: glass, background
GW-Depth	1,200	indoor and outdoor glass wall scenes; classes: glass walls, background

Within that comparison, MJU-Glass is distinctive in five explicitly stated ways. It is collected by a physical service robot rather than merely assembled from existing sources; it is built around RGB-D pairs from a realistic robot sensing setup; it uses two robot-mounted viewpoints at different heights and angles; it targets real-world service-robot navigation scenarios rather than only static scenes or limited glass-wall cases; and it includes diverse transparency, reflectivity, and environmental disturbance conditions. The authors also contrast it with RGB-D GSD, describing that resource as a curated ensemble dataset that is more diverse overall, whereas MJU-Glass is more specifically tailored to real-world service-robot glass segmentation (Lin et al., 3 Aug 2025).

This positioning matters because many glass-segmentation benchmarks have been organized around generic RGB imagery, mirror detection, transparent-object segmentation, or curated multi-source data. MJU-Glass instead binds together a robot platform, RGB-D sensing, binary glass-surface masks, and boundary-sensitive evaluation, making it a benchmark centered on deployment-relevant robotic perception rather than solely on category-level visual recognition.

6. Scope boundaries, unresolved details, and common confusions

Several important limitations are explicit or strongly implied in the paper. The exact train/validation/test split of the 4,900 images is not specified. The annotation protocol is not specified. Exact scene-category and object-type breakdowns are not specified. The depth acquisition resolution is reported inconsistently as $0.46 \text{ m/s}$7 in one place and $0.46 \text{ m/s}$8 in another, although aligned benchmark depth is used at $0.46 \text{ m/s}$9. The dataset is restricted to binary semantic segmentation and does not report instance-level or multi-class transparent-object annotations (Lin et al., 3 Aug 2025).

The paper nevertheless presents MJU-Glass as a public platform for evaluating segmentation models and states that it is publicly available at https://github.com/weduake/MJU_Glass. It also suggests future extensions to additional modalities such as thermal or polarization, and to other tasks such as transparent-environment object detection (Lin et al., 3 Aug 2025).

A recurrent source of confusion is that several recent glass-related papers do not use MJU-Glass at all. The L+GNet study evaluates only on GDD, Trans10k-Stuff, GSD, and HSO (Ojala et al., 4 Mar 2026). MGNet evaluates only on GSD, PMD, and Trans10k (Xu et al., 2024). The RGB-thermal segmentation paper introduces a different 5,551-pair RGB-T dataset rather than MJU-Glass (Huo et al., 2022). GlassNICOL is a separate dataset for glassware detection and subclass classification in human-robot bartending scenarios (Gajdošech et al., 6 Mar 2025). MonoGlass3D introduces a real-world dataset for monocular 3D glass detection with plane parameters and depth maps, but it is not named MJU-Glass (Zhang et al., 6 Sep 2025). The SciGlass-derived binary classification study concerns whether an inorganic composition will form a glass and is unrelated to visual glass-surface segmentation (Carvalho et al., 16 Mar 2026).

In the current literature, MJU-Glass is therefore best understood as a robotics-oriented RGB-D benchmark for binary glass-surface segmentation. Its defining contribution is not merely adding more glass images, but coupling real robot viewpoints, two-view RGB-D capture, challenging transparent-surface conditions, and boundary-aware evaluation in a form directly relevant to service-robot safety and perception (Lin et al., 3 Aug 2025).