Our paper entitled "Geometry-Aligned Sketch Representation Learning via Line-Jet Masked Autoencoding" has been accepted by IEEE International Conference on Multimedia and Expo 2026 (ICME 2026).

Title: Geometry-Aligned Sketch Representation Learning via Line-Jet Masked Autoencoding  

Author: Bowen Wang, Haoran Yu, Jin Wang, Xiaoming Zhao,Weifeng Liu^*

Abstract: Sketch representation learning is becoming increasingly important for a wide range of applications, such as drawing-based psychological analysis. However, mainstream image representation learning methods are not tailored to sparse, geometry-dominated sketches, and tend to bias the learned representations toward texture and appearance rather than stroke geometry. To this end, we propose Line-Jet Masked Autoencoder (LJ-MAE), which reconstructs a geometric field deterministically computed from Gaussian scale-space derivatives. This field encodes stroke saliency with gradient magnitude, unoriented direction with a half-turn periodic encoding, and stroke bending with bounded level-line curvature. It enables geometry-aligned masked reconstruction over the sketch manifold without relying on vector strokes, skeletonization, or graph construction. Extensive experiments and ablation studies on TU-Berlin, TUB-Scene, and the domain-specific HTP dataset demonstrate the superiority of our proposed method.