MiniShift: 3D Anomaly Detection Benchmark

• MiniShift is a high-resolution 3D anomaly detection dataset designed for industrial inspection and subtle defect identification.
• It features fixed dense point clouds with 500,000 points and employs a controllable synthetic anomaly pipeline to simulate micro-defects.
• The benchmark introduces a three-level difficulty protocol and pairs with Simple3D to evaluate precise anomaly localization.

Searching arXiv for the MiniShift paper and closely related benchmark context. MiniShift is a high-resolution 3D anomaly detection dataset for industrial inspection with subtle geometric defects, introduced together with the Simple3D framework in "Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects" (Cheng et al., 10 Jul 2025). It comprises 2,577 point clouds, each with 500,000 points, and anomalies occupying less than $1\%$ of the total; it is positioned as the first benchmark that explicitly targets fine-scale, subtle 3D defects at industrial resolutions rather than the relatively coarse and visually obvious anomalies common in prior 3D anomaly detection datasets. The dataset is built on industrial CAD/scan data, uses a controllable synthetic anomaly pipeline, and is intended to force methods to handle high-resolution point clouds and tiny anomaly regions.

1. Definition, motivation, and scope

MiniShift was created to address two limitations identified in existing 3D anomaly detection benchmarks. First, prevailing methods typically downsample point clouds to approximately $1\text{k}$ points, operate at group level, and then interpolate anomaly scores back to the dense cloud; the stated consequence is that this interpolation destroys spatial granularity, making millimeter-scale defects undetectable (Cheng et al., 10 Jul 2025). Second, although some existing datasets provide relatively high-resolution scans, their anomalies are described as large and visually obvious, unlike industrial micro-scratches, slight dents, or tiny protrusions.

The dataset therefore targets an inspection regime in which scanners routinely capture hundreds of thousands to millions of points per object and real failures often stem from subtle, small-area defects that occupy far less than $1\%$ of the surface. MiniShift was created to provide a benchmark where such subtle anomalies actually exist and matter, to force methods to handle high-resolution point clouds and tiny anomaly regions, and to offer a controlled, scalable synthesis pipeline so difficulty and defect type can be systematically varied.

Its object vocabulary is inherited from MulSen-AD normal point clouds and spans 12 industrial categories: Capsule, Cube, Spring pad, Screw, Screen, Piggy, Nut, Flat pad, Plastic cylinder, Button cell, Toothbrush, and Light. The result is a dataset whose defining properties are fixed dense sampling, low anomaly occupancy, and explicit difficulty structuring.

2. Dataset construction and AG-GAS synthesis

MiniShift is built on top of MulSen-AD normal point clouds and uses Anchor-Guided Geometric Anomaly Synthesis (AG-GAS), an automated framework that synthesizes geometric defects by selecting two anchor points on the surface, computing a geodesic path between them, expanding that path into a local region, and applying a smooth geometric distortion along the region (Cheng et al., 10 Jul 2025).

For each normal mesh or point cloud, dense sampling is performed to obtain

$$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$

This fixed resolution is a core design decision: every sample is represented at 500,000 points.

AG-GAS begins by randomly selecting two anchor points,

$$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$

which determine the location and span of the synthetic anomaly. A $k$-NN graph is then constructed over the point cloud, with Euclidean edge weights

$$w(\boldsymbol{e}_{ij}) = \|\boldsymbol{p}_i-\boldsymbol{p}_j\|_2.$$

Using Dijkstra's algorithm, the shortest geodesic path between the anchors is computed as

$$\pi^* = \mathop{\arg\min}\limits_{\pi \in \Pi(\boldsymbol{p}_s, \boldsymbol{p}_e)} \sum_{\boldsymbol{e}_{ij} \in \pi} w(\boldsymbol{e}_{ij}),$$

where $\Pi(\boldsymbol{p}_s, \boldsymbol{p}_e)$ denotes all paths between the anchors. This path acts as the central axis of the defect.

Given a control radius $r$, AG-GAS defines a deformation region $1\text{k}$0 around the path: $1\text{k}$1 The set $1\text{k}$2 directly serves as the point-wise ground-truth defect mask. Points in this region are displaced along a mean normal direction to create a smooth protrusion or depression, with displacement largest along the central path and tapering to zero at the region boundary. The synthesis is parameterized to generate four defect types: Areal, Striate, Scratch, and Sphere.

These four defect types correspond to different geometric regimes. Areal defects are local bulges or depressions of moderate area. Striate defects are elongated ridge or valley defects with limited width but noticeable length. Scratch defects are thin, long, line-like defects with extremely small coverage. Sphere defects are very small localized bumps or pits.

3. Difficulty protocol, splits, and annotation

MiniShift introduces a three-level difficulty protocol—Easy, Medium, and Hard—defined by normalized geometric parameters rather than by informal visual labels (Cheng et al., 10 Jul 2025). The normalization uses the bounding-box diagonal

$1\text{k}$3

the geodesic distance $1\text{k}$4 between anchors, the region radius $1\text{k}$5, and the maximum distortion amplitude $1\text{k}$6, with

$1\text{k}$7

Difficulty varies by defect type. For Areal and Striate defects, difficulty is primarily controlled by distortion depth $1\text{k}$8, with Easy corresponding to larger deformations and Hard to micro-scale distortions such as $1\text{k}$9. For Scratch, difficulty is governed by length $1\%$0 and radius $1\%$1, so shorter and narrower scratches are harder. For Sphere, difficulty is controlled by radius $1\%$2, with smaller spheres being harder. The protocol is intended to ground difficulty in geometry and perceptual saliency while keeping anomaly area below $1\%$3 of the surface.

MiniShift contains 1,137 normal training point clouds and 1,440 test point clouds, of which 1,200 are anomalous and 240 are normal. Each category has 130 test samples: 10 normal, 30 Areal, 30 Striate, 30 Scratch, and 30 Sphere. The test set is logically partitioned into MiniShift-Easy, MiniShift-Medium, MiniShift-Hard, and the merged MiniShift-ALL.

Ground truth is point-wise because AG-GAS explicitly defines the deformation region $1\%$4. Each point is labeled normal or anomalous, and each sample also carries a sample-level label, a defect-type label, and an implicit difficulty label. The recommended evaluation protocol is to train only on the normal training set, then report object-wise and point-wise AUROC separately for Easy, Medium, Hard, and MiniShift-ALL.

For AUROC computation, the paper defines

$1\%$5

Object-wise AUROC treats each point cloud as one instance with a scalar anomaly score, whereas point-wise AUROC treats each point as a binary instance with a score.

4. Position among 3D anomaly detection benchmarks

MiniShift is presented in direct comparison with MVTec 3D-AD, Real3D-AD, Anomaly-ShapeNet, and MulSen-AD (Cheng et al., 10 Jul 2025). Its distinguishing features are its fixed 500,000-point resolution, its average anomaly coverage ratio of $1\%$6, and its explicit Easy/Medium/Hard protocol.

Dataset	Resolution (pts)	Anomaly ratio
MVTec 3D-AD	26,532	1.8%
Real3D-AD	169,022	2.7%
Anomaly-ShapeNet	58,604	2.4%
MulSen-AD	29,818	4.2%
MiniShift	500,000	0.99%

In the same comparison, MiniShift has 12 categories and 2,577 total point clouds. The paper identifies four key differences relative to prior benchmarks: highest spatial resolution, smallest anomaly footprint, explicit difficulty scaling, and diverse synthetic defects with geometric realism. The claim is not merely that the point clouds are denser, but that the benchmark changes the effective detection regime: previous benchmarks do not sufficiently stress high-resolution or micro-defect detection capability.

A common misconception is that any dataset with high-resolution scans already tests subtle-defect detection. MiniShift is explicitly framed against that assumption: the paper distinguishes between scan resolution and effective anomaly difficulty, emphasizing that visually obvious anomalies remain easier even when the underlying scans are relatively dense. A plausible implication is that benchmark design in 3D anomaly detection depends jointly on resolution, anomaly occupancy, and the geometric subtlety of the defect process rather than on point count alone.

5. Associated method: Simple3D

MiniShift is accompanied by Simple3D, an unsupervised, prototype-based anomaly detector designed for high-resolution point clouds and subtle defects (Cheng et al., 10 Jul 2025). Its design uses handcrafted geometric descriptors rather than heavy CNN or ViT backbones, specifically FPFH features, Multi-scale Neighborhood Descriptors (MSND), Local Feature Spatial Aggregation (LFSA), and nearest-neighbor deviation from a prototype set of normal features.

For each point $1\%$7, Simple3D computes multi-scale $1\%$8-nearest-neighbor neighborhoods,

$1\%$9

typically with $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$0, applies a local feature operator $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$1 instantiated with CUDA-accelerated FPFH from Open3D, and concatenates the resulting descriptors: $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$2 These descriptors are intended to encode local geometric patterns at several neighborhood radii.

LFSA aggregates these MSND features around sampled anchor points. If $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$3 points are sampled from $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$4, the aggregated feature for sampled point $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$5 is

$$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$6

where $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$7 denotes the set of MSND features in the local neighborhood of $$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$8. This averaging is described as smoothing local noise, expanding receptive field, and improving discriminability of subtle local changes.

All aggregated features from normal training clouds form a prototype set

$$\boldsymbol{P} \in \mathbb{R}^{n \times 3},\quad n = 500{,}000.$$9

For a test feature $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$0, the nearest prototype is found and the anomaly score is the feature-space distance to that prototype: $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$1 Point-wise scores are interpolated back to the full cloud, and the object-wise anomaly score is

$$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$2

Simple3D is described as training-free in the deep learning sense, since it does not rely on backpropagation over a deep backbone. Its purpose within the MiniShift framework is methodological as well as empirical: it is intended to show that efficient feature aggregation can preserve fine local detail while avoiding the quadratic complexity associated with Transformer-based architectures over many point groups.

6. Empirical results, use cases, and limitations

On MiniShift-ALL, the reported AUROC values are $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$3 for PatchCore-FP, $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$4 for GLFM, and $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$5 for Simple3D in object-wise/point-wise evaluation, respectively (Cheng et al., 10 Jul 2025). By difficulty subset, Simple3D reports $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$6 on Easy, $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$7 on Medium, and $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$8 on Hard. The paper states that performance degrades for all methods as difficulty increases and that some baselines approach chance level on Hard, while Simple3D maintains its strongest relative margin in point-wise localization.

Appendix-level results further indicate that Simple3D achieves the best object-wise or point-wise AUROC in 10 of 12 categories on MiniShift-ALL, with average $$\boldsymbol{p}_s, \boldsymbol{p}_e \in \boldsymbol{P},$$9, compared with $k$0 for PatchCore-FP and $k$1 for GLFM. Ablations report that both MSND and LFSA contribute gains, that increasing the detection resolution $k$2 from 256 to 8192 monotonically improves performance, and that increasing aggregated neighborhood size $k$3 from 8 to 128 also improves performance. Runtime analysis shows that above approximately 4096 sampled points, Simple3D surpasses prior state-of-the-art methods while remaining above 20 fps, which the paper characterizes as real-time industrial applicability.

MiniShift is intended for benchmarking 3D anomaly detection on realistic industrial resolutions, evaluating robustness to micro-defects and small anomaly ratios, and studying tradeoffs between resolution, speed, and accuracy. The recommended usage is an unsupervised anomaly detection setting: train on the normal training set only, test on all four defect types across Easy, Medium, and Hard subsets, and report both object-wise and point-wise AUROC.

The dataset resources listed in the paper include a project page, a GitHub repository for Simple3D implementation and evaluation scripts, and a HuggingFace dataset release. The paper also frames several limitations. MiniShift contains synthetic anomalies only, not defects captured from real defective parts; it covers 12 categories and 4 defect types; it focuses on surface geometry rather than internal defects or multi-sensor anomalies; and it provides binary anomaly masks rather than defect hierarchies or severity estimates. The authors state that they plan to leverage the scalable anomaly generation pipeline to build a larger-scale, more diverse 3D anomaly detection dataset and to facilitate foundation-model-like 3D anomaly detectors.

These limitations also clarify the benchmark's evidentiary scope. MiniShift is not a real-defect corpus, not a multimodal benchmark, and not a defect-taxonomy dataset. Its central contribution is narrower and more specific: a controlled high-resolution benchmark for subtle geometric anomaly detection, coupled with a baseline that argues for the importance of high-resolution data and effective feature aggregation in practical 3D anomaly detection.