Researchers introduce XShapeEnc, a training-free method to encode 2D geometric shapes for neural networks without requiring model fine-tuning.

XShapeEnc is a general-purpose encoding strategy that converts arbitrary 2D spatial geometric shapes into compact representations for deep learning tasks

The method decomposes shapes into normalized geometry within a unit disk and pose vectors, enabling better neural network compatibility

The encoding exhibits five favorable properties including invertibility, adaptivity, and frequency richness without requiring any training

Addresses the limitation of traditional positional encoding, which works well for 1D sequences but struggles with 2D spatial geometric data

Technical report published on arXiv (arXiv:2604.07522v1) in the computer vision field