Continual 3D Anomaly Detection (C3D-AD)

• The paper introduces a unified framework (C3D-AD) that mitigates catastrophic forgetting by learning 3D point cloud anomalies incrementally without full data replay.
• It employs Kernel Attention with random features (KAL) for efficient feature extraction and a learnable advisor (KAA) to update task-specific knowledge continuously.
• Representation stability is reinforced using Reconstruction with Parameter Perturbation (RPP), achieving AUROC scores of 66.4%, 83.1%, and 63.4% on benchmark datasets.

Searching arXiv for the target paper and closely related works mentioned in the provided data. First, I’ll locate the C3D-AD paper by its arXiv ID, then I’ll look up the backbone/related methods explicitly named in the provided material so any citations remain grounded. Continual 3D Anomaly Detection (C3D-AD) is an end-to-end continual learning framework for 3D anomaly detection in point clouds, introduced to address two limitations identified for existing methods: class-specific training and the inability to learn from emerging classes over time. The framework is designed for a class-incremental setting in which tasks arrive sequentially, each task contains normal samples from disjoint categories, prior full datasets are not revisited, and testing is cumulative over all encountered tasks. Its architecture combines Kernel Attention with random feature Layer (KAL) for generalized feature extraction, Kernel Attention with learnable Advisor (KAA) for continual task-wise knowledge updating in the encoder and decoder, and Reconstruction with Parameter Perturbation (RPP) for representation rehearsal under bounded adversarial perturbation. On three public datasets, the reported average performance is 66.4%, 83.1%, and 63.4% AUROC on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD, respectively (Lu et al., 2 Aug 2025).

1. Problem setting and continual learning protocol

C3D-AD is formulated for 3D Anomaly Detection (AD) on high-precision industrial products represented as 3D point clouds. The stated goal is to detect anomalies or defects while avoiding the inefficiency of training one model per product category. The motivating difficulty is that a unified model trained sequentially on emerging classes without revisiting all previous data suffers catastrophic forgetting (Lu et al., 2 Aug 2025).

The continual protocol is class-incremental. Training data arrives as a sequence of disjoint tasks,

$$\mathcal{P}_{\textrm{train}}=\mathcal{P}^1_{\textrm{train}} \cup \mathcal{P}^2_{\textrm{train}} \cup \cdots \cup \mathcal{P}^T_{\textrm{train}},$$

with

$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$

At time $t$, only the current task’s normal data and a small memory, if any, can be accessed; prior full datasets are not revisited. Testing is cumulative:

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$

Evaluation is conducted with object-level AUROC (O-AUROC), and the paper reports mean O-AUROC over categories together with average ranking, where lower is better. The continual splits are four tasks for Real3D-AD, four tasks for Anomaly-ShapeNet, and three tasks for MulSen-AD. A plausible implication is that the framework is intended to model deployment conditions in which new product categories appear progressively while detection must remain valid on all previously encountered categories.

2. Architectural composition and end-to-end pipeline

The pipeline begins from an input point cloud $P \in \mathbb{R}^{N \times 3}$. Points are grouped by Furthest Point Sampling (FPS) into $n$ local neighborhoods around sampled centers, with an adaptive radius $r$ per center to handle scale differences. For a center $\bar{\mathbf{p}}_i$, its neighborhood is defined as

$$\mathcal{N}_r(\bar{\mathbf{p}}_i)=\{\mathbf{p}_j \in \mathbf{P} \mid \|\bar{\mathbf{p}}_i-\mathbf{p}_j\|_2 \le r\}.$$

The radius is set adaptively from inter-center distances and a scaling factor $\eta$, which the paper states is used for scale normalization (Lu et al., 2 Aug 2025).

Feature tokens are then extracted through KAL, which maps queries and keys into a unified kernel, or Hilbert, space using random feature approximations. These tokens pass through an encoder-decoder architecture employing KAA in both encoder and decoder attention blocks. The advisor $$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$0 is updated per task to learn new category-specific information while reducing redundant old information. Finally, RPP imposes a rehearsal constraint by reconstructing tokens and enforcing consistency between the model’s current outputs and its predicted future outputs under worst-case bounded parameter perturbations.

Anomaly detection is performed by comparing feature tokens and reconstructed tokens. The paper states that anomalies are detected “according to the anomaly score by comparing the differences between feature tokens and reconstruction tokens.” At inference time, an object-level anomaly score $$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$1 is obtained from token reconstruction discrepancy, and AUROC is computed across objects. No hard threshold is fixed because AUROC is threshold-free.

3. Kernelized feature extraction with KAL

KAL is the feature extraction mechanism introduced to extract generalized local features from diverse product types of different tasks efficiently. It replaces standard quadratic self-attention with linear-time kernel attention using a positive random feature map $$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$2, thereby providing what the paper describes as a normalized feature space across diverse categories (Lu et al., 2 Aug 2025).

For reference, standard attention is written as

$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$3

Kernel attention replaces the exponential similarity by a kernel:

$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$4

where

$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$5

Using a FAVOR$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$6-style rearrangement, the attention output becomes

$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$7

This reduces computational complexity from quadratic to linear in the number of tokens or groups.

The positive random feature mapping is given as

$$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$8

with $$\mathcal{P}^i_{\textrm{train}} \cap \mathcal{P}^j_{\textrm{train}}=\emptyset \quad (i\ne j).$$9 i.i.d. and small $t$0. The paper attributes two roles to this construction: positivity and stabilized attention computation, and embedding features into a common kernel space that promotes generalized local feature extraction across categories and tasks. This suggests that KAL addresses continual learning not only through efficiency but also through representation alignment under inter-task distribution shift.

4. Continual knowledge updating with the learnable advisor

KAA is the mechanism by which C3D-AD updates knowledge continually in both encoder and decoder blocks. Ignoring the denominator in Eq. (2), the attention output can be written as a linear transform of the kernelized query:

$t$1

where $t$2 is the learnable advisor storing and updating task-by-task knowledge (Lu et al., 2 Aug 2025).

The advisor is optimized through the objective

$t$3

Here, $t$4 aligns the advisor’s projected key $t$5 with the value $t$6 through a directional alignment term. The gradient given in the paper is

$t$7

With learning rate $t$8, the update rule is

$t$9

The paper also provides an additive-subtractive decomposition. Let

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$0

and

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$1

Then

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$2

The paper explains the first term as reducing redundant information from previous tasks and the second as learning new information. The task-wise output is

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$3

Within the stated design, KAA functions as a learnable memory with linear complexity and without a large memory bank. This suggests that the framework locates continual adaptation primarily in the advisor parameters rather than in explicit replay of raw data.

5. Representation rehearsal with parameter perturbation

RPP is introduced to preserve representation consistency across tasks. Its central idea is to constrain the model’s hypothesis over time by aligning current outputs with future outputs under worst-case bounded perturbation. The paper first writes

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$4

Since $$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$5 is unknown at time $$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$6, the future parameter is approximated by $$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$7, leading to the optimization

$$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$8

The perturbation $$\mathcal{P}^1_{\textrm{test}} \subset \mathcal{P}^2_{\textrm{test}} \subset \cdots \subset \mathcal{P}^T_{\textrm{test}}=\mathcal{P}_{\textrm{test}}.$$9 is obtained by gradient ascent in the constrained parameter space, producing an adversarial rehearsal signal intended to stabilize representations against future drift (Lu et al., 2 Aug 2025).

The paper further states a generalization error bound:

$P \in \mathbb{R}^{N \times 3}$0

This bound is used to justify keeping $P \in \mathbb{R}^{N \times 3}$1 small in practice. The text explicitly notes that large $P \in \mathbb{R}^{N \times 3}$2 increases generalization error and that Table 4 shows AUROC varies with $P \in \mathbb{R}^{N \times 3}$3.

The reconstruction objective is described as minimizing the discrepancy between input feature tokens and reconstructed tokens. The main text does not fix a particular geometric distance such as Chamfer distance or EMD. A common misconception would be to treat C3D-AD as a point-level geometric reconstruction method with a specified 3D distance; the provided description does not support that reading. The stated emphasis is token-space reconstruction combined with KAA and RPP.

6. Training procedure, datasets, and empirical behavior

The training procedure begins with pre-training a feature extractor, PointMAE with KAL, on ModelNet40, which the paper writes as “ModelNet408K.” KAA advisors are initialized in encoder and decoder blocks, with $P \in \mathbb{R}^{N \times 3}$4, $P \in \mathbb{R}^{N \times 3}$5, the random feature dimension $P \in \mathbb{R}^{N \times 3}$6 by default, and scaling factor $P \in \mathbb{R}^{N \times 3}$7 for adaptive radius. RPP’s perturbation budget $P \in \mathbb{R}^{N \times 3}$8 is scanned over $P \in \mathbb{R}^{N \times 3}$9. Optimization uses AdamW with initial learning rate $n$0, reduced to $n$1 after 800 epochs, and the architecture uses 4 stacked encoder-decoder blocks (Lu et al., 2 Aug 2025).

For each task, only normal samples of the current categories are used. The model performs tokenization with FPS and adaptive neighborhoods, extracts token features via KAL, applies KAA in each attention block with

$n$2

updates the advisor per batch via Eq. (4) and optionally Eq. (5), minimizes reconstruction discrepancy, and jointly minimizes the RPP objective under bounded perturbations. At inference, the model reconstructs token sequences and computes anomaly scores from token reconstruction differences.

The datasets and continual splits reported in the paper are as follows.

Dataset	Data description	Continual split
Real3D-AD	1,254 high-resolution samples, 12 categories; training per category has only 4 normal samples	4 tasks
Anomaly-ShapeNet	1,600 synthetic samples across 40 categories; high inter-class diversity	4 tasks
MulSen-AD	2,035 samples from 15 categories; training 1,391 normal; test 150 normal + 494 anomalous	3 tasks

The reported mean AUROC values are 66.4% for Real3D-AD, 83.1% for Anomaly-ShapeNet, and 63.4% for MulSen-AD. Continual per-task O-AUROC is reported as 0.805, 0.658, 0.666, and 0.664 on Real3D-AD; 0.854, 0.848, 0.862, and 0.831 on Anomaly-ShapeNet; and 0.698, 0.650, and 0.634 on MulSen-AD. In multi-class, non-continual evaluation on Anomaly-ShapeNet across 40 categories, C3D-AD achieves Mean O-AUROC 0.846 and best average ranking 2.78.

The baseline comparisons listed in the paper include Continual-Reg3D-AD, Continual-PatchCore (FPFH), and Continual-MC3D-AD. The paper states that C3D-AD outperforms these continual baselines and attributes the gains to unified kernel feature space through KAL, advisor memory through KAA, and representation rehearsal through RPP. Ablation results indicate that removing KAL, KAA, or RPP degrades performance, with the full model obtaining 0.664 on Real3D-AD, 0.831 on Anomaly-ShapeNet, and 0.634 on MulSen-AD. The paper further states that KAL and RPP contribute most to continual performance, while KAA yields gains by balancing new and old information.

7. Complexity, resource trade-offs, and stated limitations

C3D-AD is described as having linear-time kernel attention in both KAL and KAA via the random features $n$3, with inference time and memory scaling linearly with the number of groups or tokens (Lu et al., 2 Aug 2025). The advisor dimension $n$4 controls a performance-resource trade-off:

$n$5	AUROC / time / memory
10	0.802 / 0.281 s / 4.431 GB
1000	0.842 / 0.393 s / 5.680 GB
5000	0.819 / 1.109 s / 9.400 GB

The reported behavior is that increasing $n$6 improves AUROC up to a point while linearly increasing memory and time; too large an $n$7 does not guarantee better accuracy. The paper also states that good performance is obtained with approximately 4096 groups or tokens. Overhead is mainly in the advisor $n$8 and the random feature map $n$9, and no replay buffer of raw samples is needed.

Several limitations are explicitly noted. The perturbation radius $r$0 should be small because large $r$1 increases generalization error. The hyperparameters $r$2 and $r$3 affect the balance between reducing redundant old information and learning new information. Some categories in Anomaly-ShapeNet are not top-1, and individual classes may favor specialized methods such as diffusion-based R3D-AD or geometry-aware MC3D-AD. The method assumes normal-only training per task. The paper concludes that “further research is needed to explore how to effectively constrain the advisor to achieve better performance in continual 3D AD,” which points to advisor regularization or capacity control as an open direction.

In reproducibility terms, the paper states that code would be available upon acceptance, and reports an implementation environment of PyTorch 1.13.0, CUDA 11.7, and NVIDIA A100-PCIE-40GB. Seeds are not specified, while implementation details are said to be included in supplemental materials.