Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation

Conference: ICCV 2025 arXiv: 2505.24431 Code: https://github.com/ZZZBBBZZZ/PASDF Area: 3D Vision / Anomaly Detection / Point Cloud Keywords: 3D anomaly detection, signed distance function, pose alignment, anomaly repair, point cloud

TL;DR

This paper proposes PASDF, a framework that employs a pose-aware signed distance function (SDF) for continuous geometric representation, unifying 3D anomaly detection and repair tasks, achieving state-of-the-art performance on Real3D-AD and Anomaly-ShapeNet.

Background & Motivation

1. Importance of 3D anomaly detection: In manufacturing quality control, robotic manipulation, and related domains, even minor 3D anomalies—missing features, deformations, or shape irregularities—can cause complete component failure, necessitating robust 3D anomaly detection.

2. Limitations of existing methods:

   • Voxel-based methods: Discretization causes loss of fine geometric details and suffers from cubic memory growth
   • Point cloud methods: Sparse sampling leads to density inconsistency and incomplete surface coverage
   • Projection-based methods: Information loss in occluded regions and view-dependent distortions
   • All these approaches are inherently discrete representations that introduce quantization artifacts, hindering fine-grained anomaly localization

3. Need to bridge detection and repair: In 3D printing and advanced manufacturing, anomaly detection is only the first step; in-situ repair is equally critical. Conventional methods rely on indirect feature mappings and cannot provide explicit shape reconstruction to guide repair. Even reconstruction-based methods (e.g., IMRNet, R3D-AD) fail to generate continuous, high-fidelity repair templates due to their reliance on discrete point cloud representations.

4. Core motivation: Leveraging the continuity and smoothness of SDFs to bridge anomaly detection and repair, while decoupling pose to address detection under arbitrary orientations.

Method

Overall Architecture

PASDF consists of three core stages:

1. Pose Alignment Module (PAM): Aligns input point clouds to a canonical coordinate frame, eliminating the effect of pose variation
2. SDF Network: Learns a continuous signed distance function representation that implicitly captures object shape
3. Anomaly Scoring Module: Computes anomaly scores based on the deviation of SDF values for test samples

Theoretically, anomaly detection is formulated as evaluating the likelihood that a test sample conforms to the normal shape distribution. Pose invariance is achieved by integrating over the SE(3) group, approximated in practice by aligning to a canonical pose.

Key Designs

1. Pose Alignment Module (PAM)

PAM adopts a coarse-to-fine two-stage registration strategy:

• Coarse alignment: Voxel downsamples the raw point cloud, extracts FPFH (Fast Point Feature Histogram) features, and applies RANSAC for global coarse registration
• Fine alignment: Refines the coarse alignment using ICP (Iterative Closest Point)
• Iterative optimization: Introduces a Chamfer Distance-driven feedback mechanism that dynamically adjusts the loss threshold τ to avoid local minima. The accumulated transformation matrix is updated iteratively as: \(T^{(k)} = T_{icp}^{(k)} \cdot T_{ransac}^{(k)} \cdot T^{(k-1)}\)

Key PAM parameters: loss threshold τ = 0.016, increment Δτ = 0.001, maximum iterations K = 10.

2. SDF Network

The aligned point cloud is represented via a neural network-parameterized SDF:

• Query points are sampled from three regions: surface points (10k), interior bounding-box points (10k), and unit volume points (3k), totaling 23,000 points
• Query coordinates are encoded using sinusoidal positional encoding \(\gamma(\mathbf{x}_i)\)
• Network architecture: 8-layer MLP with weight normalization, ReLU activations, dropout (0.2) in intermediate layers, and a skip connection at the fourth layer
• SDF predictions are trained with a truncated L1 loss

3. Anomaly Score Computation

• Point-level anomaly score: \(A(\mathbf{x}_j) = |f_\theta(\mathbf{x}_j)|\), i.e., the absolute value of the SDF output
• Object-level anomaly score: Mean of the top-K (K = 1000) highest point-level anomaly scores

4. Anomaly Repair

The trained SDF network implicitly encodes the "normal" shape manifold: 1. The anomalous input is aligned to the canonical pose via PAM 2. The Marching Cubes algorithm extracts the zero-level set of the SDF 3. A point cloud is sampled from the resulting triangular mesh as the repaired output

Loss & Training

• Truncated L1 loss: \(\mathcal{L}_{SDF} = \frac{1}{N_q} \sum_{i=1}^{N_q} |\text{clamp}(\hat{s}_i, -d_{max}, d_{max}) - s_i|\)
• Truncation distance \(d_{max} = 0.1\)
• Training for 2000 epochs with learning rate \(1\times10^{-5}\)
• Data preprocessing: coordinates normalized to \([0,1]^3\); non-manifold detection and Poisson reconstruction ensure watertightness

Key Experimental Results

Main Results

Datasets: - Real3D-AD: A high-resolution real-world dataset with 12 categories, 4 normal training samples and 100 test samples per category - Anomaly-ShapeNet: A synthetic dataset with 40 categories and 1600+ samples

Results on Real3D-AD

Method	O-AUROC ↑	P-AUROC ↑
BTF(Raw)	0.635	0.571
BTF(FPFH)	0.603	0.733
M3DM(PointMAE)	0.552	0.637
PatchCore(FPFH+Raw)	0.682	0.680
RegAD	0.704	0.705
IMRNet	0.725	-
Group3AD	0.751	-
PASDF (Ours)	0.802	0.745

PASDF surpasses Group3AD by 5.1% in O-AUROC, with particularly strong performance on Seahorse (1.000), Car (0.959), and Fish (0.989).

Results on Anomaly-ShapeNet

Method	O-AUROC Mean ↑	P-AUROC Mean ↑
BTF(Raw)	0.493	0.550
M3DM	0.552	0.616
CPMF	0.559	-
RegAD	0.572	0.668
IMRNet	0.661	0.650
R3D-AD	0.749	-
PASDF (Ours)	0.900	0.897

PASDF achieves the best O-AUROC on 37 out of 40 categories and attains perfect scores of 1.000 on multiple categories.

Ablation Study

Effect of PAM on Various Baselines (Anomaly-ShapeNet)

Method	PAM	O-AUROC ↑	P-AUROC ↑
BTF(FPFH)	✗	0.528	0.628
BTF(FPFH)	✓	0.579	0.683
PatchCore(FPFH)	✗	0.568	0.580
PatchCore(FPFH)	✓	0.814	0.867
PatchCore(PointMAE)	✗	0.562	0.577
PatchCore(PointMAE)	✓	0.626	0.681
PASDF (Full)	✓	0.900	0.897

Incorporating PAM into PatchCore(FPFH) yields gains of +24.6% in O-AUROC and +28.7% in P-AUROC, a remarkably substantial improvement.

Component Ablation (Anomaly-ShapeNet)

Method	O-AUROC ↑	P-AUROC ↑
w/o RANSAC	0.711	0.739
w/o ICP	0.727	0.836
w/o Iterative Optimization	0.871	0.884
w/o Positional Encoding	0.887	0.783
PASDF (Full)	0.900	0.897

Anomaly Repair Quality Evaluation

Method	Real3D-AD CD ↓	Real3D-AD EMD ↓	Anomaly-ShapeNet CD ↓	Anomaly-ShapeNet EMD ↓
w/o PE	0.0255	0.0133	0.0575	0.0276
with PE	0.0203	0.0110	0.0445	0.0228

Key Findings

1. Pose alignment is critical: Removing RANSAC causes O-AUROC to drop sharply from 0.900 to 0.711, confirming the necessity of global coarse registration.
2. Dual role of positional encoding: Removing PE degrades P-AUROC by 11.4% (0.897→0.783) and simultaneously deteriorates repair quality.
3. Generalizability of PAM: PAM functions as a plug-and-play module that consistently boosts multiple baseline methods.
4. Advantage of continuous representation: PASDF outperforms all competing methods on 37/40 categories, whereas other methods exhibit unstable performance across categories.

Highlights & Insights

1. Unified framework: PASDF is the first to unify 3D anomaly detection and repair under a continuous SDF representation, with both tasks sharing the same learned shape prior.
2. Pose–shape disentanglement: PAM explicitly decouples pose from shape, allowing the SDF network to focus on intrinsic shape variations.
3. Continuous vs. discrete: Replacing discrete voxel/point cloud/projection representations with a continuous SDF eliminates quantization artifacts and preserves fine geometric details.
4. Plug-and-play PAM: Beyond serving PASDF, PAM substantially improves the performance of other methods, demonstrating broad generalizability.

Limitations & Future Work

1. Computational cost: PAM registration can be expensive in cases with challenging initial poses; learning-based or hierarchical registration approaches may be explored.
2. Single-class assumption: The current formulation assumes a single normal object category; extending to multi-class detection would improve practical applicability.
3. Sensitivity to input quality: Performance is affected by input point cloud quality, motivating enhanced robustness to noise and outliers.
4. Absence of contextual information: Integrating scene context could improve detection in complex real-world environments.
5. Training efficiency: The 2000-epoch training schedule is lengthy; more efficient training strategies warrant investigation.

Related Work & Insights

• DeepSDF: The neural network parameterization of SDFs originates from DeepSDF; PASDF creatively adapts this idea for anomaly detection.
• ICP + RANSAC: The combination of these classical registration methods remains highly effective within PAM.
• Marching Cubes: This classical algorithm is employed in the repair stage to extract the zero-level set from the SDF.
• Insight: Continuous representations (implicit functions) hold a natural advantage over discrete counterparts in tasks requiring fine-grained geometric perception.

Rating

• Novelty: ⭐⭐⭐⭐ — The unified SDF-based framework for 3D anomaly detection and repair is a novel and well-motivated contribution
• Technical Quality: ⭐⭐⭐⭐ — The method is thoroughly designed, ablations are comprehensive, and the generalizability of PAM is convincingly validated
• Experimental Thoroughness: ⭐⭐⭐⭐ — Two datasets, multiple baselines, detailed ablations, and qualitative results
• Value: ⭐⭐⭐⭐ — The unified detection-and-repair framework has direct applicability in manufacturing scenarios
• Overall: ⭐⭐⭐⭐ (8/10)

Related Papers

• [ICCV 2025] PASDF: Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation
• [ICCV 2025] SiM3D: Single-Instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark
• [ICCV 2025] RayletDF: Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians
• [ICCV 2025] Momentum-GS: Momentum Gaussian Self-Distillation for High-Quality Large Scene Reconstruction
• [ICCV 2025] DAP-MAE: Domain-Adaptive Point Cloud Masked Autoencoder for Effective Cross-Domain Learning