ShotcreteDepth: Industrial Depth Perception Dataset

• ShotcreteDepth is a bi-modal dataset combining stereo RGB images and LiDAR to tackle depth perception in challenging shotcrete construction environments.
• It captures adverse conditions like high dust, turbidity, and poor illumination, providing realistic test cases for stereo matching, depth completion, and estimation.
• The dataset uses precision synchronization, custom sensor housings, and a detailed annotation pipeline to address sensor degradation and support robust robotic navigation.

Searching arXiv for papers on ShotcreteDepth and closely related shotcrete depth perception work. ShotcreteDepth is a bi-modal dataset from the construction domain that captures both an active shotcreting process and general construction environments. It comprises stereo RGB imagery and LiDAR point clouds acquired under harsh real-world conditions, including high turbidity and poor illumination, and was introduced to support stereo matching, depth completion, and depth estimation under conditions that closely reflect the operational complexities found in industrial settings (Gregorek et al., 22 Jun 2026).

1. Motivation and operational setting

ShotcreteDepth was created for robotic depth perception in shotcrete construction, where autonomous systems require reliable perception for navigation, mapping, obstacle avoidance, safety, material-deposition monitoring, and quality control. The underlying premise is that existing construction datasets rarely capture the specific failure modes that arise during shotcreting. Shotcreting sites are described as especially difficult because the process generates dense airborne dust and often takes place in confined spaces such as tunnels or mines, with illumination coming only from artificial lighting. These conditions introduce strong turbidity, poor visibility, and sensor degradation, making both camera-based and laser-based depth sensing unreliable in different ways (Gregorek et al., 22 Jun 2026).

The dataset therefore addresses a narrowly defined but practically important regime: real-world data from a niche industrial environment in which stereo matching, depth completion, and monocular depth estimation can be studied under adverse conditions rather than under clean laboratory assumptions. The dataset spans scenes before shotcreting, during active shotcreting, after shotcreting, and broader construction-site environments around the process. This diversity is significant because the appearance and sensor quality change substantially depending on whether shotcrete dust is present, whether surfaces are freshly sprayed, and how severely the environment is obscured or brightly lit (Gregorek et al., 22 Jun 2026).

A central feature of the collection protocol is that the workspace was sealed off from the outside, so illumination came exclusively from artificial lights. Shotcrete particles accumulated rapidly in the air, causing high turbidity and poor visibility. The paper notes that stereo imagery becomes harder to match under these conditions, while LiDAR can be partially or completely blinded, can produce noisy measurements, and can even detect dust clouds as spurious returns. This setup makes ShotcreteDepth a dataset not merely about geometric depth, but about robust depth perception under sensor failure modes intrinsic to active spraying (Gregorek et al., 22 Jun 2026).

2. Dataset composition and sensing configuration

ShotcreteDepth contains 11,252 temporally synchronized samples, of which 220 are manually annotated for evaluation. The sensing setup combines a Roboception rc_visard 160c stereo camera and a Velodyne PUCK LiDAR. The camera provides RGB images at $1280 \times 960$ resolution, while its internal stereo engine outputs disparity and confidence maps at $640 \times 480$. LiDAR depth measurements are projected into the image plane, yielding an aligned depth modality for evaluation and cross-modal learning (Gregorek et al., 22 Jun 2026).

The hardware arrangement is also part of the dataset’s design logic. To protect the sensors in the spray environment, the camera and LiDAR were mounted inside a custom 3D-printed dust-proof housing, with the LiDAR positioned approximately centered above the stereo camera. The authors additionally tested a Neuvition Titan M1 solid-state LiDAR, but report that it was not suitable for this environment because it failed to return useful depth in the turbid conditions (Gregorek et al., 22 Jun 2026).

Component	Specification	Role
Stereo camera	Roboception rc_visard 160c; RGB at $1280 \times 960$; disparity/confidence at $640 \times 480$	Image and stereo modality
LiDAR	Velodyne PUCK	Sparse depth modality projected to image plane
Collection scale	11,252 synchronized samples; 220 annotated	Full corpus and evaluation subset

The paper’s parameter table further specifies that the stereo camera has a $61^\circ$ horizontal field of view and a $48^\circ$ vertical field of view, while the LiDAR has a $30^\circ$ vertical field of view. LiDAR depth is projected at the same image resolution as the camera for evaluation. This multi-modal alignment is central to the dataset’s intended use in supervised evaluation, cross-modal transfer, and sparse-to-dense reconstruction tasks (Gregorek et al., 22 Jun 2026).

3. Calibration, synchronization, and annotation pipeline

ShotcreteDepth places unusual emphasis on synchronization and calibration because active shotcreting produces rapidly changing scene structure and transient dust clouds. Intrinsic camera calibration was performed using Roboception’s built-in calibration tool, and the extrinsic camera–LiDAR relationship was estimated using MATLAB’s Camera and LiDAR calibration tools. For timing, LiDAR timestamps were made accurate via PPS and NMEA signals from a GPS module, while the camera clock was synchronized using PTP to the acquisition computer, which itself was synchronized to GPS time (Gregorek et al., 22 Jun 2026).

Because the modalities operate at different rates—RGB at 25 Hz, stereo disparity at 3 Hz, and LiDAR at 10 Hz—the dataset pairs RGB images and disparity maps with the temporally closest LiDAR point cloud and retains only complete synchronized sets containing all modalities. This decision is technically consequential: under active spraying, even small temporal offsets would confound correspondence between dust, geometry, and sparse depth, thereby degrading both training and evaluation (Gregorek et al., 22 Jun 2026).

Only the 220 evaluation samples are annotated in detail. To make this feasible, the authors release a lightweight point-cloud annotation tool designed for time-efficient labeling of LiDAR point clouds. The tool allows each point to be labeled as “kept by user” or “removed by user,” and it also supports corresponding algorithmic labels, namely “kept by algorithm” and “removed by algorithm,” for automatic filtering. Its interface contains two synchronized views: a 3D visualization of the point cloud and a projection of the point cloud over the left RGB image. It also allows images to be flagged for evaluation. The purpose of this tool is to label LiDAR dust clouds and other unreliable measurements that should be excluded from ground truth (Gregorek et al., 22 Jun 2026).

An additional curation step removes occluded points. The paper follows a sliding-window occlusion-filtering approach similar to prior work: projected depth points are scanned with a window, and points farther than the nearest depth in that window by more than a threshold are removed. Because the LiDAR has fewer scan lines than in typical automotive datasets, the authors use a $10 \times 50$ pixel window and a 0.5 m distance threshold. Manually marked removals are preserved so that valid points near dust clouds are not mistakenly discarded. This processing yields the filtered evaluation depth against which algorithms are measured (Gregorek et al., 22 Jun 2026).

4. Task formulation and evaluation protocol

ShotcreteDepth explicitly supports three research tasks: stereo matching, depth completion, and depth estimation. The common evaluation setup runs all methods at $640 \times 480$ inference resolution and then upscales predictions to the original $1280 \times 960$ image size for metric computation. The annotated and filtered LiDAR point clouds serve as ground truth. For depth completion, the authors sample 500 uniformly distributed sparse depth points from stereo matching and use them as input (Gregorek et al., 22 Jun 2026).

For affine-invariant depth estimation methods, the same sparse points are used to fit scale $640 \times 480$0 and shift $640 \times 480$1 by least squares: $640 \times 480$2 Here $640 \times 480$3 denotes the model prediction and $640 \times 480$4 the sparse depth reference. This alignment is necessary because some predictors estimate depth only up to an unknown affine transform. For conversion between depth and disparity, the dataset uses the standard stereo relation

$640 \times 480$5

with baseline determined by the stereo rig and focal length supplied by camera calibration (Gregorek et al., 22 Jun 2026).

The benchmark metrics are task-specific. For depth estimation, the paper reports absolute relative error,

$640 \times 480$6

and the threshold accuracy $640 \times 480$7, defined as the percentage of pixels satisfying

$640 \times 480$8

For depth completion, it reports

$640 \times 480$9

and

$1280 \times 960$0

Because sparse ground truth can hide boundary artifacts, the dataset also uses Pseudo Depth Boundary Error accuracy and completeness. For stereo matching, it reports end-point error in disparity space and D1, the fraction of pixels whose disparity error exceeds 3 pixels and 5% of the ground truth. Runtime and model size are also included, measured on an Nvidia GeForce 4090 (Gregorek et al., 22 Jun 2026).

This protocol makes ShotcreteDepth relevant to both algorithmic robustness and deployment efficiency. A plausible implication is that the dataset is designed not only for offline benchmarking but also for examining the trade-off between recovery quality and computational practicality under industrial constraints.

5. Benchmarks and empirical behavior

The benchmark study includes several representative methods in each task category. Stereo methods include RAFT-Stereo, FoundationStereo, Stereo Anywhere, and the built-in Semi-Global Matching implementation of the rc_visard camera. Depth completion methods include Marigold-DC, Marigold-SSD, and VPP4DC. Depth estimation methods include Depth Anything v3, Marigold-E2E, and MoGe-2, with and without scale-shift alignment where appropriate (Gregorek et al., 22 Jun 2026).

The reported experimental pattern is consistent across tasks: the dataset is genuinely challenging. The neural stereo methods achieve stronger and denser results than the classical onboard SGM, especially under darkness and overexposure, though at the cost of larger model sizes and slower runtime. The general conclusion drawn from the benchmarks is that depth perception is possible in shotcrete environments, but there is a clear trade-off between speed and accuracy: lightweight systems can run near real time, while large foundation models are more accurate but significantly slower (Gregorek et al., 22 Jun 2026).

These findings are significant because they show that ShotcreteDepth is not merely a passive repository of difficult imagery. It functions as a stress test for perception systems in an environment where the dominant failure modes are modality-specific and physically induced. Stereo can degrade because turbidity destroys matchability, while LiDAR can degrade because dust introduces missing data and spurious returns. This suggests that the dataset’s main scientific value lies in exposing complementary weaknesses across sensing modalities, rather than in providing idealized supervision (Gregorek et al., 22 Jun 2026).

The authors also make a forward-looking observation that stereo cameras and LiDAR have complementary strengths and weaknesses, so sensor fusion may eventually deliver the best solution for shotcrete environments. This is presented cautiously: doing so well would require an additional, more complete source of depth supervision or evaluation (Gregorek et al., 22 Jun 2026).

6. Limitations, common misconceptions, and relation to other “depth” concepts

A common misconception would be to treat ShotcreteDepth as a fully labeled dense-depth benchmark. The paper explicitly states otherwise. The dataset is not fully labeled; only the 220 evaluation samples are annotated in detail, and the ground truth is derived from noisy real sensors that must be filtered and annotated. The labeled subset is therefore curated rather than exhaustive, and the dataset does not provide dense pixel-perfect supervision across the full collection (Gregorek et al., 22 Jun 2026).

Another limitation is sensor dependence in the evaluation reference itself. The ground truth relies on completion of the annotation and filtering pipeline over LiDAR returns, including the removal of dust-cloud artifacts and occluded points. This means that the reference depth is operationally useful but not equivalent to an ideal external metrology source. The failure of the tested Neuvition Titan M1 solid-state LiDAR in the same environment further underscores that the sensing problem is materially conditioned by turbidity and hardware characteristics (Gregorek et al., 22 Jun 2026).

The term “depth” in ShotcreteDepth also benefits from disambiguation. In adjacent shotcrete and concrete literature, “depth” can refer to very different quantities. In shotcrete 3D printing quality control, the closest analogue is filament thickness or height, estimated from segmentation masks via a distance transform, where the paper states that filament thickness is twice the height of the distance transform results (Mawas et al., 26 Nov 2025). In concrete non-destructive evaluation with ground-penetrating radar, depth can refer to embedded rebar cover depth inferred by matching extracted hyperbola outlines to a theoretical database (Xiang et al., 2020), or to dielectric-aware depth prediction used in pose-aware 3D GPR migration, where depth is linked to two-way travel time through the medium’s effective wave velocity (Feng et al., 2020). By contrast, ShotcreteDepth denotes a dataset for robotic scene-depth perception in active shotcreting environments, not a single scalar thickness, cover, or penetration-depth variable (Gregorek et al., 22 Jun 2026).

Taken together, these distinctions clarify the dataset’s place within a broader technical vocabulary. ShotcreteDepth is best understood as infrastructure for evaluating robust geometric perception under adverse construction conditions. Its principal contribution is not a new depth variable, but a benchmark regime in which depth sensing itself becomes the central research problem (Gregorek et al., 22 Jun 2026).