Nature3D-AD: Geometry-Aware Feature Learning for Natural-Growth 3D Anomaly Detection

3D anomaly detection is a crucial process in agricultural production chains, playing a vital role in automated processing and ensuring food safety. However, existing 3D anomaly detection datasets are primarily designed for industrial scenarios with standardized geometric shapes. In contrast, agricultural products originate from natural growth, exhibiting significant variations in shape, size, and texture. Consequently, the definition of "normal" is far less explicitly constrained than that of industrial components, rendering current industrial datasets and detection methods inadequate for this application. To bridge this gap, we make contributions from both data and method perspectives. On the data side, we introduce NaturalGrowth, the first dataset dedicated to agricultural point cloud anomaly detection. On the method side, we propose Nature3D-AD, which incorporates two novel modules. Anomaly-Sensitive Positional Encoding (ASPE) integrates local curvature and density information to provide geometry-aware positional representations. Geometry-Aware Attention (GAA) injects geometric biases in the spatial domain and captures global structural patterns through a spectral FFT branch. Extensive experiments on NaturalGrowth, Anomaly-ShapeNet, and Real3D-AD demonstrate that Nature3D-AD achieves state-of-the-art performance. The dataset and code will be released.