CASL: Curvature-Augmented Self-supervised Learning for 3D Anomaly Detection

AI summary

Overview

Research area: 3D computer vision, specifically point cloud anomaly detection and self-supervised point cloud representation learning.

Technical level: Intermediate. Readers need familiarity with point cloud self-supervised learning (masked reconstruction, Point-MAE-style pipelines), U-Net encoder-decoder architectures, and standard anomaly detection evaluation metrics (AUROC).

Scope: The paper proposes CASL, a curvature-augmented self-supervised reconstruction framework built on a Minkowski-convolution U-Net, which learns general-purpose 3D representations that transfer to anomaly detection via a simple pseudo-anomaly classification fine-tuning step.

What This Paper Is About

Dedicated 3D anomaly detection methods detect anomalies well but cannot transfer to other 3D tasks, while general-purpose self-supervised point cloud models transfer broadly but perform poorly at anomaly detection when fine-tuned in the standard "pre-training and fine-tuning" way. The authors trace this gap to "geometric shortcuts" — reconstruction from coordinate space to coordinate space lets models rely on low-level positional cues rather than meaningful geometry. Their goal is a single, generalizable model that achieves leading anomaly detection through ordinary fine-tuning, without any anomaly-specific detection machinery.

Key Contributions

1. Demonstrating the power of curvature alone. The authors show that a non-learning method using only per-point curvature as the anomaly score outperforms several classical self-supervised models and task-specific anomaly detection methods on benchmark datasets, establishing curvature as a highly informative cue for 3D anomalies.

2. A curvature-augmented reconstruction framework (CASL). Built on a U-Net architecture, CASL masks out all point coordinates and reconstructs them using multi-scale curvature prompts injected into the decoder, forcing the network to learn geometry from curvature rather than from overlapping coordinate semantics.

3. Anomaly detection without anomaly-specific design. Detection is achieved by appending a binary classification head and fine-tuning with pseudo-anomalies generated by perturbing patches along their normal direction (following PO3AD), with point-level anomaly scores derived from the classification logits.

4. Leading detection results plus generalization. CASL improves average Object-level AUROC by 5.6% on Real3D-AD and 4.8% on Anomaly-ShapeNet over the prior best, and also transfers to point cloud classification and part segmentation.

Main Findings

• Curvature correlates with anomalies: In the authors' visualizations, points at the edges of abnormal regions show significantly higher curvature than surrounding normal regions, and anomalous point boundaries often exhibit sharp increases in curvature.

• Curvature beats several learned baselines: On Real3D-AD, the non-learning curvature method reaches 0.723 O-AUROC and 0.729 P-AUROC, exceeding task-specific methods such as M3DM (0.552 O-AUROC), CPMF (0.586 O-AUROC), and Reg3D-AD (0.704 O-AUROC), as well as self-supervised models Point-BERT, Point-MAE, and PointGPT reported in Figure 1.

• CASL sets new state of the art on Real3D-AD: CASL achieves 0.823 average O-AUROC and 0.882 average P-AUROC across the 12 categories, surpassing the second-best approach (ISMP at 0.767 O-AUROC and 0.836 P-AUROC) by 5.6% in O-AUROC and 4.6% in P-AUROC.

• CASL sets new state of the art on Anomaly-ShapeNet: Averaged over 40 categories, CASL reaches 0.887 O-AUROC and 0.899 P-AUROC, surpassing the second-best PO3AD (0.839 O-AUROC, 0.898 P-AUROC) by 4.8% in average O-AUROC.

• Full masking works best: The ablation comparing four configurations shows that self-reconstruction from coordinates to themselves (no masking, no curvature) gives the lowest reconstruction loss (0.008) but poor detection (0.751 O-AUROC, 0.649 P-AUROC). Adding 60% masking improves this to 0.776/0.711 (reconstruction loss 0.012). Curvature alone without masking reaches 0.804/0.878 (loss 0.019), and the full method with 100% masking plus curvature prompts reaches 0.823/0.882 (loss 0.016). Performance is best at a 100% masking ratio.

• Strong transfer to classification with little pre-training data: On ScanObjectNN, CASL achieves 92.08% on OBJ-BG, 91.05% on OBJ-ONLY, and 86.81% on PB-T50-RS. It leads on the first two variants but trails SFR (87.8) on PB-T50-RS. Notably, CASL is pre-trained on only 832 samples, versus roughly 50k samples of ShapeNet used by prior approaches.

• Competitive part segmentation: On ShapeNetPart, CASL obtains mIoU_c of 84.1 and mIoU_I of 86.2, matching PointGPT-S (84.1 / 86.2).

Methodology in Plain English

The authors first argue that the standard reconstruction approach — hide some point coordinates, predict them from the visible ones — creates a "geometric shortcut" because the model can lean on position information rather than learning meaningful shape. Their fix is to remove coordinates entirely from the input side and hand the model curvature instead.

Computing curvature. For each point, they take its k nearest neighbors, compute a 3×3 covariance matrix of the centered neighborhood, decompose it into eigenvalues ordered λ1 ≤ λ2 ≤ λ3, and define curvature as (λ1 + λ2 + λ3) / λ1. This is a purely geometric quantity computed directly from coordinates, with no learning.

The architecture. Instead of a Transformer — which the authors note is inefficient for the hundreds of thousands of points typical in this setting, due to patch-based operations and O(n²) complexity — they use a classical U-Net built on Minkowski convolution, which is well-suited to high-resolution point clouds. It has three parts:

• A curvature encoder that embeds point-wise curvature with an MLP and passes it through three encoder blocks, each downsampling by 2 (kernel and stride of 2) followed by a convolution block of four stride-1 layers with residual connections.
• A coordinate encoder that takes randomly initialized features for the N masked points and passes them through four encoder blocks, producing a tensor of shape N4 × 256.
• A fusion decoder that upsamples with Minkowski Transpose Convolution (kernel and stride 2), concatenates the result with the curvature prompts at the matching resolution, and applies convolution blocks. The final N × 96 output is reduced by MLPs to N × 32, concatenated with the original curvature features, and mapped by a second set of MLPs into 3D coordinate space.

Training objective. Because reconstructed objects typically contain over 100k points, Chamfer Distance and Earth Mover's Distance are computationally infeasible, so the authors combine ℓ1 and ℓ2 losses between the reconstructed and ground truth point clouds.

Pre-training data. The pre-training set is built from the training splits of Real3D-AD (12 categories, 4 normal samples each for training, 100 test instances per category) and Anomaly-ShapeNet (a synthetic ShapeNet-derived set with 1,600 point cloud samples across 40 categories, 4 normal training instances each). This yields 208 distinct normal samples; four augmented variants per sample give 832 normal samples total. The model is trained for 300 epochs with Adam, learning rate 0.001, decayed by a factor of 0.5 every 10 epochs.

Fine-tuning. Pseudo-anomalies are synthesized by dividing the point cloud into patches and displacing all points in randomly selected patches along the surface normal or its opposite by a controlled distance. A binary classification head is appended after the first MLPs, so detection follows an ordinary fine-tuning pipeline. Because reconstruction is not needed at fine-tuning time, the coordinate encoder can use embeddings taken directly from coordinates rather than random initialization.

Scoring. The point-level anomaly score is φ(x_i) = −log(x_i^0 / (x_i^1 + ε)), the log ratio of normal to anomalous predicted probabilities. The sample-level score averages the top-k point scores, where k = ⌈r · N⌉ and r is an aggregation rate.

Why This Matters

Impact on research. The paper challenges the assumption that 3D anomaly detection requires task-specific architectures. It offers evidence that a general-purpose self-supervised model, trained with a curvature-based semantic target, can match or beat dedicated detectors while remaining transferable to classification and segmentation. The "geometric shortcut" diagnosis — that coordinate-to-coordinate reconstruction causes fine-grained representational collapse — is a transferable insight for other masked-reconstruction pipelines, not just anomaly detection.

Real-world applications.

• Industrial quality control and defect inspection on manufactured parts, where anomalous protrusions, indentations, or sharp edges manifest as curvature changes.
• Reverse engineering and CAD model validation, where scanned geometry must be checked against specifications.
• Robotic grasping and manipulation, where detecting local surface irregularities affects contact reliability.
• 3D scanning and digitization pipelines, where corrupted or malformed scans need to be flagged before downstream processing.

Industry relevance. Real3D-AD is collected from real-world objects, which grounds the results in manufacturing-like conditions. The method requires only normal samples for pre-training and fine-tuning, avoiding the high cost of collecting labeled defects. The curvature signal is cheap to compute and the Minkowski U-Net backbone is more efficient than Transformers for the high point counts typical in industrial scanning. The efficient use of only 832 pre-training samples is relevant where large 3D corpora are unavailable.

Future Directions

• Testing curvature augmentation on larger and more diverse pre-training corpora. The authors note that CASL pre-trains on 832 samples versus roughly 50k for prior methods and still matches or exceeds them; whether the gap on the PB-T50-RS variant of ScanObjectNN (86.81 vs. SFR's 87.8) closes with more data is untested here.

• Extending beyond curvature as the only geometric prior. The paper establishes curvature as highly informative; whether other intrinsic surface descriptors (or combinations) offer further gains, and whether they reduce the geometric shortcut even more, is left open.

• Broadening the set of downstream 3D tasks evaluated. Only classification and part segmentation are tested here alongside anomaly detection; other tasks such as semantic segmentation or completion are not reported.

• Understanding and quantifying the geometric shortcut more directly. The paper diagnoses the problem through detection performance and reconstruction loss rather than through a direct measurement of representational collapse, so a more principled diagnostic remains an open question.

Target Audience

Researchers and engineers working on 3D point cloud representation learning, industrial visual inspection, or self-supervised learning for structured data. It is most useful to readers who already understand the standard masked-reconstruction fine-tuning paradigm and want to see how the choice of reconstruction target — coordinates versus intrinsic geometry — changes what a model actually learns. Practitioners seeking a deployable anomaly detection pipeline that requires only normal training samples and no task-specific detection module will also find it directly relevant.

Authors’ abstract

Deep learning-based 3D anomaly detection methods have demonstrated significant potential in industrial manufacturing. However, many approaches are specifically designed for anomaly detection tasks, which limits their generalizability to other 3D understanding tasks. In contrast, self-supervised point cloud models aim for general-purpose representation learning, yet our investigation reveals that these classical models are suboptimal at anomaly detection under the unified fine-tuning paradigm. This motivates us to develop a more generalizable 3D model that can effectively detect anomalies without relying on task-specific designs. Interestingly, we find that using only the curvature of each point as its anomaly score already outperforms several classical self-supervised and dedicated anomaly detection models, highlighting the critical role of curvature in 3D anomaly detection. In this paper, we propose a Curvature-Augmented Self-supervised Learning (CASL) framework based on a reconstruction paradigm. Built upon the classical U-Net architecture, our approach introduces multi-scale curvature prompts to guide the decoder in predicting the spatial coordinates of each point. Without relying on any dedicated anomaly detection mechanisms, it achieves leading detection performance through straightforward anomaly classification fine-tuning. Moreover, the learned representations generalize well to standard 3D understanding tasks such as point cloud classification. The code is available at https://github.com/zyh16143998882/CASL.