PIAD: Pose and Illumination agnostic Anomaly Detection

November 11, 2025 · View on GitHub

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. Source code and data will be made publicly available upon paper acceptance.

Overview

The codebase has 4 main components:

A PyTorch-based optimizer to refinement camera pose.
A CUDA kernel that supports backpropagation of camera poses.
A pre-trained network for camera pose initialization based on EfficientLoFTR.
An anomaly detection module.

Setup

Local Setup

First, you need to install the required packages. You can create a new conda environment (python=3.8) and then run:

pip install -r requirements.txt

And then download the pre-trained EfficientNet-B4 weights from the PAD GitHub repository and put them in corresponding file location:

cd retrieval
gdown https://drive.google.com/uc\?id\=1CVFXJf203tDLKDNF2lx1FYNxWPNrZyt-
unzip model.zip

the file format is like this:

retrieval
 └ model

Additionally, you will need to download the pre-trained weights for EfficientLoFTR. You can find the download link in their GitHub repository. (or from here) The required file is eloftr_outdoor.ckpt, which should be placed in the ./EfficientLoFTR/weights/ directory.

the file format is like this:

EfficientLoFTR
 └ weights 
    └ eloftr_outdoor.ckpt

Datasets

You can download the MAD dataset from the PAD GitHub repository, and it should be placed in ./data/LEGO-3D/.
You can download our dataset from here, and it should be placed in ./data/MIP/ (for synthetic data) or ./data/Colmap/ (for real data). Alternatively, you can use the example data Spring and 05Joint, which has already been placed in ./data/MIP/and ./data/Colmap/.

Running (Pose Estimation)

To run the optimizer, simply use

python pose_estimation.py --config <path to config.txt> --class_name <class_name>

The pose_estimation.py script first generates a set of reflection images and then trains 3DGS; these two steps are executed only during the initial run. Finally, it estimates the camera pose for each query image, renders the reference images, and saves the results to ./output.