PADFormer detects anomalies from sparse views without 3D reconstruction

Detecting anomalies in industrial or medical images becomes far trickier when objects appear under arbitrary viewpoints. Traditional pose-agnostic anomaly detection (PAD) methods have relied on complex 3D reconstruction, which demands extensive multi-view data and heavy computation. A new paper accepted as an oral at ECCV 2026 introduces PADFormer, a purely image-space method that sidesteps these costs. The model, built on a Vision Transformer, learns to reconstruct anomaly-free versions of query images while preserving the original pose. Its key innovation is adapting cross-view masked reconstruction to train exclusively on normal data, combined with dynamic patch selection and spatial alignment mechanisms. This lets the model learn effectively from just a few sparse reference views, even when viewpoints vary significantly.

At inference, PADFormer runs multiple forward passes with different masking patterns, generating an ensemble of anomaly-free reconstructions that comprehensively cover the query image. Anomalies are then flagged by comparing these reconstructions against the original. The approach achieves state-of-the-art results on the PAD benchmark while maintaining comparable performance on classic few-shot anomaly detection (FSAD) tasks. Crucially, it eliminates the need for 3D reconstruction entirely, offering superior efficiency and generalization. This makes PADFormer a practical option for real-world scenarios where only sparse, arbitrary-view images are available, such as factory floors, remote inspection, or medical imaging.