WUDataset: Dual Benchmark for UDA & Localization

• WUDataset in the WUDA setting comprises controlled source–target pairs with weak bounding box labels and styled target variants to analyze domain shift in semantic segmentation.
• WUDataset for inertial localization provides synchronized, high-frequency IMU and magnetometer data with precise ground truth to benchmark uncertainty-aware pedestrian tracking.
• Despite sharing a name, the two WUDataset artifacts differ in modality, annotation, and application, underscoring the importance of contextual clarification in research.

Searching arXiv for “WUDataset” and closely related entries to verify current usage and disambiguate the term. “WUDataset” is an ambiguous label in the arXiv literature. It denotes at least two unrelated research artifacts: a controlled suite of source–target domain pairs for weakly supervised unsupervised domain adaptation in semantic segmentation, introduced with “WUDA: Unsupervised Domain Adaptation Based on Weak Source Domain Labels” (Liu et al., 2022), and a large-scale inertial dataset for pedestrian localization introduced alongside ReNiL in “ReNiL: Relative Neural Inertial Locator with Any-Scale Bayesian Inference” (Wu et al., 8 Aug 2025). In the first usage, WUDataset is not a single monolithic dataset but a set of domain-shifted image collections with weak source labels and no target labels; in the second, it is a synchronized multimodal motion corpus comprising IMU, magnetometer, and ground-truth trajectories for indoor and outdoor pedestrian motion. The shared name masks substantial differences in modality, annotation regime, evaluation protocol, and intended research use.

1. Terminological scope and disambiguation

In the semantic-segmentation setting, WUDataset was introduced to study a newly defined task, unsupervised domain adaptation based on weak source domain labels, or WUDA (Liu et al., 2022). The dataset is organized around source–target pairs rather than a single homogeneous corpus. Its purpose is to analyze how weak supervision in the source domain, specifically bounding boxes, interacts with varying degrees of domain shift. The construction includes “real-to-real” and “synthetic-to-real” pairs, as well as stylized target-domain variants that span a spectrum of representation shifts.

In the inertial-localization setting, WUDataset was developed at Wuhan University to support any-scale, demand-driven inertial localization methods under realistic motion conditions (Wu et al., 8 Aug 2025). Here the emphasis is on high-rate synchronized sensing, diverse carrying modes, multiple motion intensities, and high-fidelity ground truth from Vicon or LiDAR-SLAM. The dataset is explicitly designed for benchmarking probabilistic inertial localization and uncertainty-aware trajectory estimation.

This suggests that any reference to “WUDataset” in technical writing requires immediate contextual qualification. Without such qualification, the term is underspecified.

2. WUDataset in weakly supervised unsupervised domain adaptation

The WUDA-associated WUDataset comprises source–target domain pairs for urban-scene semantic segmentation under weak source supervision (Liu et al., 2022). The primary pairs are GTAV $\rightarrow$ Cityscapes and BDD100k $\rightarrow$ Cityscapes. The GTAV source contains approximately 24,966 images extracted from the game engine, while Cityscapes provides 2,975 training images, 500 validation images, and 1,525 test images. In the BDD100k $\rightarrow$ Cityscapes pair, the source consists of 70,000 training and 10,000 validation images at resolution $1280\times720$.

The source annotations are bounding boxes rather than full semantic masks. For GTAV, the authors derive boxes from segmentation masks by running a connected-component pass over each class label. For BDD100k, the official instance-level boxes are reused directly, but only for “thing” classes such as car, person, truck, bus, motorcycle, bicycle, and rider. Consequently, the dataset has markedly different weak-label coverage across the two sources: GTAV yields boxes for all 19 Cityscapes classes, whereas BDD100k provides boxes only for “thing” classes, leaving “stuff” classes effectively unboxed in the source domain.

Target annotations are absent during training. Pseudo-labels are instead generated by the learning pipelines themselves. This design is central to the task definition: weak supervision is confined to the source domain, and adaptation to the target domain remains unsupervised.

To probe domain shift systematically, the authors also create augmented Cityscapes variants by applying seven style transformations, including enhanced CycleGAN, FDA low-frequency exchange, simple color jitter, and three artistic filters—frosted glass, mural, and poster. Each stylized Cityscapes variant preserves the 2,975-image training-set size. No new annotations are added. A plausible implication is that the dataset functions not only as a benchmark for WUDA algorithms but also as a controlled environment for analyzing the sensitivity of transfer performance to progressively altered target appearance.

3. Representation shift and analytical role

A defining feature of the WUDA-associated WUDataset is its use in quantifying domain difficulty through the “representation shift” metric (Liu et al., 2022). Let the segmentation backbone produce a $c$-channel activation map

$F(x)=\{f_1(x),\ldots,f_c(x)\},$

with each channel of spatial size $h\times w$. The channel-wise global average is defined as

$$a_c(x)=\frac{1}{h\cdot w}\sum_{i=1}^{h}\sum_{j=1}^{w} f_c(x)_{i,j}.$$

For source and target image sets, empirical distributions over these channel averages are estimated, and the representation shift is

$$R(X^S,X^T)=\frac{1}{c}\sum_{c=1}^{c} W\!\left(p_{a_c}^S,p_{a_c}^T\right),$$

where $W(\cdot,\cdot)$ is the one-dimensional Wasserstein distance.

In practice, the authors train DeepLabV3 with ResNet-38 or ResNet-101 on the source set without data augmentation, freeze the feature backbone, compute channel-average histograms for source and target images, approximate the one-dimensional $\rightarrow$0, and average over channels to obtain a scalar $\rightarrow$1. They report a strong negative Pearson correlation between $\rightarrow$2 and held-out target-domain mIoU, ranging from $\rightarrow$3 to $\rightarrow$4 depending on the class set.

The analytical significance of WUDataset in this setting lies less in raw scale than in controllable heterogeneity. Because the target domain is available in multiple stylized versions with increasing shift from GTAV, the dataset supports direct examination of how feature-space divergence affects adaptation quality. The paper further notes that if $\rightarrow$5, standard UDA priors may collapse. That statement is framed as a usage note rather than a universal law, but it situates the dataset as a tool for pre-assessing transfer feasibility.

4. Annotation, preprocessing, and reproducibility in the WUDA setting

The WUDA repository hosts code and pre-built dataset splits under a structured directory layout (Liu et al., 2022). The layout includes /data/GTAV/, /data/BDD100k/, /data/Cityscapes/, /data/Cityscapes_Augmented/, /labels/, /scripts/, and /models/. GTAV is accompanied by per-image JSON files of axis-aligned boxes; BDD100k includes official box annotations; Cityscapes_Augmented stores each stylized variant in its own subfolder; helper scripts convert boxes to GrabCut masks for weakly supervised semantic segmentation; preprocessing scripts remap class IDs and resize all images to $\rightarrow$6.

The prescribed workflow begins with bash scripts/download_data.sh, which retrieves the raw images, approximately 200 GB for GTAV and BDD100k combined. Preprocessing is then run with python3 scripts/preprocess.py --input_dir data/... --output_dir data/..._processed. Training commands are provided separately for the WSSS-UDA and TDOD-WSSS frameworks.

The paper’s best-practice notes clarify how WUDataset should be used. For weak-to-strong WSSS-UDA, source boxes should cover all prominently visible instances; when large unboxed “stuff” regions are present, the authors recommend relying on the detection branch only for “thing” classes. They further state that 3 iterations of GrabCut plus 2–3 epochs of self-training improved pseudo-mask quality by approximately 7 mIoU points over a single pass. The recommendation to keep source and target crop sizes matched at $\rightarrow$7 underscores that part of the measured shift can arise from resolution mismatch rather than semantic or stylistic divergence alone.

5. WUDataset as a pedestrian inertial localization corpus

In the ReNiL paper, WUDataset denotes a large-scale inertial dataset for pedestrian motion, created to address limitations in existing public inertial datasets such as RoNIN-ds, IONet, and TLIO (Wu et al., 8 Aug 2025). The dataset spans both a Vicon-instrumented indoor laboratory of size $\rightarrow$8 and LiDAR-SLAM-reconstructed outdoor spaces including hallways, open courtyards, and corridors. It contains 120 independent trajectories totaling 40.2 km from 28 participants, with 68 indoor and 52 outdoor trajectories.

The subject pool consists of 28 healthy volunteers, 17 men and 11 women, with heights distributed over $\rightarrow$9 m. Motion is labeled at three intensity levels: slow walking, normal walking, and running, with reported speed statistics of $\rightarrow$0 m/s for slow walk, $\rightarrow$1 m/s for normal walk, and $\rightarrow$2 m/s for run. The dataset also records nine carrying modes: handheld, shoulder, arm, front pocket, back pocket, backpack, handbag, chest pocket, and swinging arm.

The sensing platform is a Google Pixel 6A smartphone. The IMU, comprising accelerometer and gyroscope, is sampled at 200 Hz, and the magnetometer is also sampled at 200 Hz. Ground truth is obtained from Vicon at 100 Hz indoors, with sub-millimeter accuracy, and from LiDAR-SLAM at 200 Hz outdoors, with centimeter accuracy. The device frame is right-handed and is rotated to the East-North-Up navigation frame via quaternion.

This WUDataset is thus a multimodal localization benchmark rather than an adaptation benchmark. Its design prioritizes synchronized sensing, trajectory fidelity, and motion diversity under realistic carrying conditions.

6. Data organization, metrics, and methodological role in inertial learning

The inertial WUDataset is organized into subject-level train, validation, and test partitions under the folder structure WUDataset/train/ val/ test/ -> subject_{ID}/ -> session_{#}/ {imu.csv, gt.csv, meta.json} (Wu et al., 8 Aug 2025). Each imu.csv contains timestamped inertial and magnetic measurements, (t, ax, ay, az, gx, gy, gz, mx, my, mz), and each gt.csv contains timestamped position and orientation ground truth, (t, x, y, q_w, q_x, q_y, q_z). The accompanying meta.json stores subject demographics, carrying mode, and motion labels.

Preprocessing includes removal of initial static segments, zero-mean acceleration bias correction, and magnetometer outlier clamping at $\rightarrow$3. IMU and ground-truth streams are interpolated to a common 200 Hz grid. The partition is a standard 5:1:4 split at subject level, and the test set is further divided into “seen” and “unseen” subsets depending on whether the subjects appear in train or validation data.

The evaluation protocol includes displacement MAE,

$\rightarrow$4

ADE per second,

$\rightarrow$5

heading error,

$\rightarrow$6

quaternion angular error,

$\rightarrow$7

and cosine similarity,

$\rightarrow$8

The recommended uncertainty-consistency reporting targets 57.7%, 95.0%, and 99.2% coverage for $\rightarrow$9, $1280\times720$0, and $1280\times720$1, respectively.

The dataset is structurally aligned with ReNiL’s modeling choices. The paper defines device-to-navigation rotation by quaternion as

$1280\times720$2

with $1280\times720$3 and $1280\times720$4, and models displacement at inertial positioning demand points through a Laplace distribution,

$1280\times720$5

A plausible implication is that WUDataset’s synchronized inertial, magnetic, and orientation signals were assembled not merely for trajectory regression, but for end-to-end learning of uncertainty-aware inertial odometry pipelines.

7. Comparative significance and common misconceptions

The two artifacts named WUDataset occupy different parts of the machine-learning landscape. The WUDA-associated version belongs to computer vision and weakly supervised domain adaptation; the ReNiL-associated version belongs to mobile sensing and inertial localization. One operates on urban-scene images with source bounding boxes and zero target labels; the other operates on time-series sensor streams with dense ground truth trajectories. One is built around domain-shift analysis and mIoU; the other around displacement accuracy, heading consistency, quaternion alignment, and uncertainty calibration (Liu et al., 2022, Wu et al., 8 Aug 2025).

A common misconception is to treat WUDataset as a single canonical benchmark. The literature cited here does not support that interpretation. In one case, WUDataset is explicitly “not a single monolithic dataset,” but a suite of domain pairs and stylized variants. In the other, it is a conventional named corpus with fixed sensing modalities, metadata, and subject-level splits. Another misconception would be to infer shared provenance or shared task structure from the shared name; no such relation is stated.

The shared label nevertheless highlights a broader pattern in contemporary benchmarking. Both datasets are designed around failure modes of existing benchmarks: insufficiently varied domain shift in WUDA, and insufficiently realistic motion diversity and uncertainty support in inertial localization. This suggests that the most important unifying characteristic of the name “WUDataset” is not modality but methodological intent: in both cases, the dataset is meant to expose weaknesses in prevailing evaluation practice and to enable more discriminating empirical analysis.