PAD: A Dataset and Benchmark for Pose-agnostic Anomaly Detection (2310.07716v1)

Abstract: Object anomaly detection is an important problem in the field of machine vision and has seen remarkable progress recently. However, two significant challenges hinder its research and application. First, existing datasets lack comprehensive visual information from various pose angles. They usually have an unrealistic assumption that the anomaly-free training dataset is pose-aligned, and the testing samples have the same pose as the training data. However, in practice, anomaly may exist in any regions on a object, the training and query samples may have different poses, calling for the study on pose-agnostic anomaly detection. Second, the absence of a consensus on experimental protocols for pose-agnostic anomaly detection leads to unfair comparisons of different methods, hindering the research on pose-agnostic anomaly detection. To address these issues, we develop Multi-pose Anomaly Detection (MAD) dataset and Pose-agnostic Anomaly Detection (PAD) benchmark, which takes the first step to address the pose-agnostic anomaly detection problem. Specifically, we build MAD using 20 complex-shaped LEGO toys including 4K views with various poses, and high-quality and diverse 3D anomalies in both simulated and real environments. Additionally, we propose a novel method OmniposeAD, trained using MAD, specifically designed for pose-agnostic anomaly detection. Through comprehensive evaluations, we demonstrate the relevance of our dataset and method. Furthermore, we provide an open-source benchmark library, including dataset and baseline methods that cover 8 anomaly detection paradigms, to facilitate future research and application in this domain. Code, data, and models are publicly available at https://github.com/EricLee0224/PAD.

• The paper introduces the MAD dataset and OmniposeAD method to address the lack of diverse and standardized benchmarks in pose-agnostic anomaly detection.
• It employs a novel NeRF-based approach, achieving pixel-level AUROC of 97.8 and image-level AUROC of 90.9 across 11,000 multi-view images.
• The study establishes a robust framework for future research and industrial quality control by enabling effective multi-view anomaly detection.

An Examination of "PAD: A Dataset and Benchmark for Pose-agnostic Anomaly Detection"

The paper in focus presents an innovative approach to pose-agnostic anomaly detection (PAD) in machine vision, introducing both a new dataset and methodology to tackle prevalent challenges within this domain. The novel contributions of this work include the creation of the Multi-pose Anomaly Detection (MAD) dataset and the introduction of the OmniposeAD method, designed specifically for this application.

Core Challenges and Contributions

The paper identifies two primary challenges in pose-agnostic anomaly detection:

1. Lack of Dataset Diversity: Existing datasets are primarily designed with a pose-aligned assumption, which does not hold in real-world scenarios where objects may present anomalies from multiple, arbitrary perspectives.
2. Performance Evaluation Protocol: There is an absence of a standardized protocol for evaluating pose-agnostic anomaly detection methods, thereby hindering fair method comparisons.

To address these, the authors present the MAD dataset, which captures over 11,000 images of 20 LEGO toys from various viewpoints, encompassing both simulated and real-world scenarios to ensure a diverse and realistic representation of object anomalies. The dataset includes comprehensive pose labels and three types of anomalies, aimed at providing a representative challenge for evaluation.

Further, the paper introduces the OmniposeAD method. This approach leverages a neural radiance field (NeRF) to encode object information from diverse viewpoints, allowing for anomaly-free reference synthesis and subsequent comparison to identify anomalies. The paper benchmarks OmniposeAD against ten state-of-the-art methods across eight different paradigms, highlighting its superior performance in detecting multi-view anomalies.

Quantitative Analysis and Results

The authors conduct extensive experiments on the MAD dataset. They highlight that OmniposeAD substantially outperforms other methods, achieving an increase in pixel-level AUROC to 97.8 and image-level AUROC to 90.9. This significant improvement suggests that OmniposeAD effectively captures both local and global features necessary for detecting anomalies across multiple viewpoints.

The paper also explores the correlation of anomaly detection performance with object attributes like shape complexity and color contrast. Notably, results indicate that most methods correlate positively with color contrast and negatively with structural complexity, signifying potential avenues for refining detection capabilities by targeting these attributes.

Implications and Future Directions

This research provides a robust framework for pose-agnostic anomaly detection, applicable to a range of industrial vision applications where viewpoint variability is a norm. The MAD dataset acts as a benchmark for future research endeavors, guiding the development of anomaly detection methods that are not constrained by pose alignment.

The findings also suggest several practical implications. For instance, deploying such a methodology in manufacturing could enhance quality control processes by detecting product anomalies that may only be visible from certain angles, thereby reducing the likelihood of defective products reaching consumers.

In future work, further exploration of more sophisticated data augmentation techniques and integrating real-time anomaly detection capabilities could enhance OmniposeAD’s applicability to dynamically changing environments. An extension of the MAD dataset to include naturally varying objects (beyond rigid LEGO toys) could also expand its relevance to broader applications.

Overall, this paper makes significant strides in addressing the gaps in pose-agnostic anomaly detection, setting a foundation for future research and development of more nuanced and versatile vision systems.