PIAD: Pose and Illumination agnostic Anomaly Detection

CVPR 2025

^1,2Kaichen Yang, ^1,2Junjie Cao, ^1,2Zeyu Bai, ^1,2Zhixun Su, ³Andrea Tagliasacchi,

¹Dalian University of Technology, Dalian, China

²Key Laboratory for Computational Mathematics and Data Intelligence of Liaoning Province, Dalian, China

³Simon Fraser University, Burnaby, Canada

Supplementary Material paper(TODO) Code Dataset

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

query image

OmniPoseAD

SplatPose

ours

OmniPoseAD

SplatPose

ours

Abstract

We introduce the Pose and Illumination agnostic Anomaly Detection (PIAD) problem, a generalization of pose-agnostic anomaly detection (PAD). Being illumination agnostic is critical, as it relaxes the assumption that training data for an object has to be acquired in the same light configuration of the query images that we want to test. Moreover, even if the object is placed within the same capture environment, being illumination agnostic implies that we can relax the assumption that the relative pose between environment light and query object has to match the one in the training data.

We introduce a new dataset to study this problem, containing both synthetic and real-world examples, propose a new baseline for PIAD, and demonstrate how our baseline provides state-of-the-art results in both PAD and PIAD, not only in the new proposed dataset, but also in existing datasets that were designed for the simpler PAD problem.

RGB&Reflectance Gaussian vs. RGB Gaussian

We present the camera pose estimation results using RGB&Reflectance Gaussian and RGB Gaussian separately. When there is inconsistent lighting between the training set and the query images, using RGB&Reflectance Gaussian is more robust.

Training data

Pose estimation by RGB Gaussian

AD Heat map

Query image

Pose estimation by RGB&Relf. Gaussian

AD Heat map

Training data

Pose estimation by RGB Gaussian

AD Heat map

Query image

Pose estimation by RGB&Relf. Gaussian

AD Heat map

Training data

Pose estimation by RGB Gaussian

AD Heat map

Query image

Pose estimation by RGB&Relf. Gaussian

AD Heat map

Dataset

Introduction

We introduce a new dataset to study the PIAD problem, containing both synthetic and real-world examples. It comprises a total of 11268 multi-view images of 30 distinct industry products, including 16 synthetic and 14 real-world products. The synthetic dataset is generated using Blender, and the real-world dataset is captured using a smart phone (Redmi K40) mounted on a gimbal. The dataset is designed to be challenging, with a variety of poses, illuminations, materials and anomaly types. The dataset will be released after acceptance.

Dataset visualization

We visualize part of our dataset to demonstrate the diversity of views, materials, and types of anomalies within the dataset.

Synthetic

Motor

Spring

Keyring

Axletree

Can

Gear

Sprockets

Picker

Box

Parts

Real

Valve

Tube

Cup

USB

Cube

PaperCup

Lighter

Filter

Wheel

Bearing

BibTeX

@inproceedings{
      yang2025piad,
      title={PIAD: Pose and Illumination agnostic Anomaly Detection},
      author={Kaichen Yang and Junjie Cao and Zeyu Bai and Zhixun Su and Andrea Tagliasacchi},
      booktitle={The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR)},
      year={2025},
    }