Chulalongkorn University Building Recognition Dataset

• CUBR is a domain-specific building recognition dataset designed for mission-oriented IoT sensor systems and decentralized training experiments.
• It comprises approximately 3200 RGB images across 32 classes, capturing varied conditions like distance, angles, illumination, and occlusions.
• The dataset supports Wireless Ad Hoc Federated Learning (WAFL) setups, demonstrating improved performance with vision transformer models on edge devices.

Searching arXiv for the cited CUBR paper and closely related WAFL/UTBR work. arXiv_search tool is not available in this interface, so I will rely on the provided arXiv metadata and cite the specified paper directly. The Chulalongkorn University Building Recognition Dataset (CUBR) is a building-recognition dataset developed for mission-oriented Internet-of-Things (IoT) sensor systems in which edge devices perform not only inference but also collaborative on-device training. It was introduced as a case study centered on Chulalongkorn University in Bangkok, Thailand, with the explicit goal of supporting Wireless Ad Hoc Federated Learning (WAFL) using vision models in a smart-campus environment (Taniguchi et al., 5 Dec 2025). In the reported formulation, CUBR serves both as a visual-recognition benchmark and as an experimental substrate for decentralized learning under realistic sensor heterogeneity, device mobility, and scene variability.

1. Origin and mission orientation

CUBR was assembled to support mission-oriented IoT sensor systems in which devices at the edge are capable of collaborative on-device training. The motivating use case is a prototypical smart-campus deployment at Chulalongkorn University, where a network of low-power cameras might recognize and index campus buildings in real time for navigation, security, or asset-management services (Taniguchi et al., 5 Dec 2025).

The dataset was positioned in relation to prior WAFL experiments using the UTokyo Building Recognition Dataset (UTBR). The underlying rationale is that WAFL with Vision Transformer (WAFL-ViT) is a mission-oriented sensor system, and therefore specific datasets should be constructed for each mission. In that framing, CUBR is specialized for Chulalongkorn University as a Thailand-based case study.

Chulalongkorn University was selected because its urban campus presents a varied architectural landscape, ranging from colonial-era facades to modern high-rises, and because tropical lighting and frequent occlusions, including trees or passing pedestrians, introduce challenges that are not captured in many existing building-recognition data collections. This suggests that CUBR was designed not merely as a generic classification corpus, but as a domain-specific benchmark for real deployment conditions.

2. Dataset composition and image acquisition

CUBR comprises 32 distinct building classes, each represented by approximately one hundred photographs, for a total of roughly 3,200 RGB images (Taniguchi et al., 5 Dec 2025). The images were collected over the span of a month using four different consumer-grade devices, each with its own sensor characteristics, resolution, and lens distortion.

During capture, the photographers varied several acquisition factors: distance, from tens of meters to just a few; viewing angle, including frontal, oblique, and side views; and illumination, including direct sunlight, overcast conditions, and dusk. The images also include partial occlusions such as foliage or utility poles. The stated purpose of this design is to emulate the variability encountered in IoT deployment, so that intra-class diversity mirrors the appearance shifts faced by edge sensors.

The following summary condenses the core dataset specification.

Aspect	Value	Notes
Building classes	32	Distinct campus buildings
Total images	Roughly 3,200 RGB images	Approximately one hundred photographs per class
Collection period	Over the span of a month	Captured using four consumer-grade devices

In context, the acquisition protocol reflects a deliberate coupling between data collection and deployment assumptions. A plausible implication is that CUBR is intended to stress both visual recognition and decentralized training robustness, rather than to isolate classification performance under tightly controlled imaging conditions.

3. Annotation, partitioning, and evaluation setup

In organizing CUBR for training and validation, each class was partitioned into a train subset and a validation subset. Table I of the paper states that each class contributes between 83 and 126 images to training and between 20 and 32 images to validation (Taniguchi et al., 5 Dec 2025). Rather than holding out a separate test set, evaluation is carried out by distributing all images uniformly in an independent and identically distributed (IID) fashion across 10 simulated edge devices.

Labels consist solely of the building ID, indexed from 0 to 31. No bounding-box annotation or pixel-wise annotation is provided. Annotation was performed manually through a simple directory-based scheme, with one directory per class, followed by sanity checks to remove blurred or duplicated frames.

These choices place CUBR squarely in the category of image-level supervised classification datasets. The absence of detection or segmentation annotations narrows its immediate applicability to whole-image recognition settings, while the IID distribution across simulated devices directly aligns the dataset with federated-learning experiments. The decision not to maintain a separate test set is methodologically significant: it ties evaluation to the simulated multi-device regime rather than to a conventional centralized benchmark split.

4. Preprocessing pipeline and baseline architecture

Prior to model ingestion, each image is resized to $224 \times 224$ pixels and normalized by subtracting the ImageNet mean and dividing by its standard deviation (Taniguchi et al., 5 Dec 2025). During training, standard data-augmentation transforms are applied on the fly: random horizontal flips, small rotations up to $\pm 15^\circ$, and color-jittering of brightness and contrast.

As a baseline for CUBR, the reported study employed the Vision Transformer architecture ViT-B/16. In the specified configuration, the model embeds non-overlapping $16 \times 16$ patches into a 768-dimensional latent space, uses 12 Transformer encoder layers with 12 self-attention heads each, and concludes with a small multi-layer perceptron (MLP) head for classification.

This baseline is technically notable because the learning protocol does not update the full transformer. In the WAFL protocol, each device fine-tunes only its MLP head on local images, leaving the bulk of the ViT parameters fixed. This suggests an efficiency-oriented design in which communication and optimization focus on the classifier head rather than the full backbone, a choice that is especially relevant for edge devices with constrained compute or bandwidth.

5. WAFL formulation and training regime

The Wireless Ad Hoc Federated Learning protocol is defined through decentralized parameter exchange among neighboring devices. After each local-training round, device $n$ exchanges its head parameters $W^n$ with its neighbors $\mathrm{adj}(n)$ and updates according to

$$W^n \leftarrow W^n + \lambda \cdot \frac{\sum_{k \in \mathrm{adj}(n)} (W^k - W^n)} {|\mathrm{adj}(n)| + 1},$$

where $\lambda \in (0,1)$ controls the weight-mixing rate (Taniguchi et al., 5 Dec 2025).

In parallel, the aggregate network-wide update is written in federated-gradient form as

$$w^{t+1} = w^t - \eta \sum_{k \in K} \nabla F_k(w^t),$$

where $w^0$ represents the global model initialization, $\pm 15^\circ$0 is the learning rate, $\pm 15^\circ$1 indexes the participating devices, and $\pm 15^\circ$2 is the local loss. The local loss is the standard cross-entropy loss,

$\pm 15^\circ$3

over true label $\pm 15^\circ$4 and predicted probability $\pm 15^\circ$5.

In the experiments, each device ran SELF, described as self-training, for up to 2000 epochs in isolation, or WAFL for up to 1500 epochs with device-to-device exchanges governed by a random-waypoint mobility simulation. The combination of local head fine-tuning, neighbor-based parameter mixing, and mobility-driven connectivity defines the operational setting in which CUBR was evaluated.

6. Reported performance and observed failure modes

Under the IID distribution, WAFL-ViT achieved an average accuracy of $\pm 15^\circ$6, outperforming SELF-ViT at $\pm 15^\circ$7 and also outperforming other WAFL variants: WAFL-ResNet-152 at $\pm 15^\circ$8, WAFL-MobileNet-V2 at $\pm 15^\circ$9, and WAFL-VGG-19-BN at $16 \times 16$0 (Taniguchi et al., 5 Dec 2025).

These results were presented as evidence for the benefits of collaborative head-parameter aggregation in a fully decentralized setting. Within the confines of the reported experiment, the comparison indicates that collaborative training across devices exceeded isolated self-training, and that the ViT-based WAFL configuration produced the highest average accuracy among the listed baselines.

A detailed confusion matrix further revealed that some adjacent classes, including labels 10 and 11, remain challenging. That observation highlights a residual difficulty in separating visually similar buildings even under the reported pipeline. A plausible implication is that additional data collection targeting hard class pairs, or broader coverage of discriminative viewpoints, could improve class separability in future versions of the dataset.

7. Availability, reuse, and scholarly positioning

The full CUBR dataset, together with scripts to reproduce the WAFL-ViT training pipeline, is publicly available at the repository specified by the authors and is released under the repository’s open-source license (Taniguchi et al., 5 Dec 2025). The original paper is identified as “University Building Recognition Dataset for mission-oriented IoT sensor systems in Thailand” by Taniguchi et al.

In scholarly terms, CUBR occupies the intersection of visual place or building recognition, edge intelligence, and decentralized federated learning. Its distinctive feature is not only that it contains campus-building imagery, but that it was constructed explicitly for a mission-oriented WAFL scenario involving edge devices, device-to-device communication, and realistic sensing heterogeneity. This suggests that its primary research value lies in enabling end-to-end study of dataset design, model adaptation, and decentralized optimization within a single application-specific benchmark.